9512.net

甜梦文库

甜梦文库

当前位置：首页 >> >> # Models of Computation for Embedded System Design

Models of Computation for Embedded System Design

Department of Electronics Politecnico di Torino C. Duca degli Abruzzi 24, Torino, Italy

lavagno@polito.it

Luciano Lavagno

Alberto Sangiovanni-Vincentelli

Department of EECS University of California Berkeley, CA 94709 USA

alberto@eecs.berkeley.edu

Cadence Berkeley Laboratories 2001 Addison St. Berkeley, California USA

ellens@cadence.com

Ellen Sentovich

September 28, 1998

In the near future, most objects of common use will contain electronics to augment their functionality, performance, and safety. Hence, time-tomarket, safety, low-cost, and reliability will have to be addressed by any system design methodology. A fundamental aspect of system design is the speci cation process. We advocate using an unambiguous formalism to represent design speci cations and design choices. This facilitates tremendously e ciency of speci cation, formal veri cation, and correct design re nement, optimization, and implementation. This formalism is often called model of computation. There are several models of computation that have been used, but there is a lack of consensus among researchers and practitioners on the \right" models to use. To the best of our knowledge, there has also been little e ort in trying to compare rigorously these models of computation. In this paper, we review current models of computation and compare them within a framework that has been recently

Abstract

1

proposed. This analysis demonstrates both the need for heterogeneity to capture the richness of the application domains, and the need for uni cation for optimization and veri cation purposes. We describe in detail our CFSM model of computation, illustrating its suitability for design of reactive embedded systems and we conclude with some general considerations about the use of models of computations in future design systems.

1 Introduction

An embedded system is a complex object containing a signi cant percentage of electronic devices (generally including at least one computer) that interacts with the real world (physical environment, human users, etc.) through sensing and actuating devices. A system is heterogeneous, i.e., is characterized by the coexistence of a large number of components of disparate type and function. For example, it may contain programmable components such as micro-processors and Digital Signal Processors, as well as analog components such as A/D and D/A converters, sensors, transmitters and receivers. In the past, the system design e ort has focused on these hardware parts, leaving the software design to be done afterwards as an implementation step. However, today more than 70% of the development cost for complex systems such as automotive electronics and communication systems is attributable to software development. This percentage is increasing constantly. The challenge posed to the semiconductor industry is to provide a new generation of programmable parts and of supporting tools to help system designers develop software faster and correctly the rst time. Today much attention is devoted to the hardware-software co-design issue, i.e., to the concurrent development of Application Speci c Integrated Circuits and standard hardware components, selection of programmable components, and development of the application software that will run on them. We believe that this approach in fact enters the design process too late to explore interesting design trade-o s. The computer-aided design process should begin at a very early stage. We believe that the real key to shortening design time and coping with complexity is to start the design process before the hardware-software partitioning. For this reason, we believe that the key problem is not so much hardware-software co-design, but the sequence consisting of specifying what the system is intended to do with no bias towards implementation, of the initial functional design, its analysis to determine whether the functional design satis es the speci cation, the mapping of this design to a candidate architecture, and the subsequent 2

1.1 Embedded System Design Today

1.2 Our Design Methodology Goals

Behavioral Libraries Capture Behavior Verify Behavior

Architecture Libraries

?

Functional Level

Verify Architecture

Capture Architecture

Map Behavior to Architecture Performance Back-Annotation Refine HW/SW ?Architecture Link to HW/SW Implementation

Verify Performance

?

Mapping Level

Link to ?Architecture Verification

Architectural Level

Figure 1: Proposed design strategy performance evaluation. It is then clear that the key aspect of system design is indeed function-architecture co-design. Our approach is a design methodology that is based on the use of formal models to describe the behavior of the system at a high level of abstraction, before a decision on its decomposition into hardware and software components is taken. Our approach also facilitates the use of existing parts. As the complexity of embedded systems increases, it is unthinkable to design new systems from scratch. Already hardware components are often standard parts that are acquired from silicon vendors, and software is often incrementally upgraded from previous versions of the same product. In the future, design re-use will be the key to pro tability and market timing. In addition, the decreasing feature size of silicon manufacturing processes will make it possible to incorporate multiple microprocessors, complex peripherals, and even sensors and actuators on the same silicon substrate, which will force system developers and IC designers/manufacturers to deal with the problem of exchanging Intellectual Property in the form of designs instead of chips. The overall design strategy that we envision is depicted in Figure 1. There are, of course, other ways to design systems in common use today. The top-down nature of the design methodology that our group has advocated throughout the years may not be agreed upon by the system design community where a mixed top-down, bottom-up approach is mostly used. In our methodology, however, we believe that this approach is captured by the presence of architectural and functional libraries that could be the result of a bottom-up assembly of basic components. We strongly emphasize that no matter how the design is carried out, a rigorous framework is necessary to reduce design iterations and to improve design quality. 3

1.3 Design Strategy

At the functional level, a behavior for the system to be implemented is selected and analyzed against a set of speci cations. The de nition of speci cation and behavior is often the subject of hot debate. For some, there is no di erence between speci cation and behavior. For some, speci cation is the I/O relation of the system to realize together with a set of constraints to satisfy and of goals to achieve, and behavior is the algorithm that realizes the function to be implemented. For others, speci cation is the algorithm itself. From a purist point of view, an algorithm is indeed the result of an implementation decision from a given set of speci cations and we prefer to stick to this view in our design methodology. For example, if we specify the function that a system has to perform as \given a nonlinear function f over the set of reals, nd x so that f(x)=0", then it is a design decision to chose the Newton-Raphson algorithm or a Gauss-Seidel nonlinear relaxation algorithm. On the other hand, for an MPEG encoder, the speci cation is the encoding of the compressed stream, and any implementation that creates it from a stream of images is \correct". In this second case, the rst step of system design has already been decided upon and the designer has no freedom to alter the conceptual design.

1.3.1 Design Conception to Design Description

1.3.2 Algorithm Design

Algorithm development is a key aspect of system design at the functional level. We believe that little has been done in this domain to help the designer to select an algorithm that satis es the speci cations. The techniques and environments for this step are often application dependent. We have experience in automotive engine control 7], where the algorithms have to have strong correctness properties due to the life critical aspects inherent in this application. In addition, the \plant" to be controlled (the combination of the engine and the drive-line) is a hybrid system consisting of continuous components (drive-line) and discrete ones (engine). To assess the properties of the algorithms, one must use control theory and sophisticated simulation techniques involving mixed di erential equations-discrete event models. The understanding of the application domains yields a design methodology that integrates the application-speci c view with general-purpose techniques that could be re-used in other domains of application. It is our strong belief that this step of system design carries the maximal leverage when combined with the design methodology proposed here.

1.3.3 Algorithm Analysis

The behavior of an algorithm is veri ed by performing a set of analysis steps. Analysis is a more general concept than simulation. For example, analysis may mean the formal proof that the algorithm selected always converges, that the computation performed satis es a set of speci cations, or that the computational complexity, measured in terms of number of operations, is bounded by 4

a polynomial in the size of the input. In the view of design re-use, parts of the overall behavior may be taken from an existing library of algorithms. Since

it is the formal model that provides the framework for algorithm analysis, it is very important to decide which mathematical model to support in a design environment.

Once the algorithm has been selected, there is an intermediate step before the selection of the architecture to support its implementation: its transformation into a set of functional components that are computationally tractable. This set of functional components have to be formally de ned to ensure that the properties of the implementation of the algorithm can be assessed. To do so, the concept of models of computation is key. Most system designs use one or more of the following models of computation: Finite State Machines, Data Flow Networks, Discrete Event Systems, and Communicating Sequential Processes. A particular model of computation has mathematical properties that can be e ciently exploited to answer questions about system behavior without carrying out expensive veri cation tasks. An important issue here is how to compare and compose di erent models of computation. Once the model(s) of computation have been selected, then we can safely proceed towards the implementation of the system by selecting the physical components (architecture) of the design.

1.4 Algorithm Implementation

1.5 Our Goals

The main goal of this paper is to review and compare the most important models of computations using a unifying theoretical framework introduced recently by Lee and Sangiovanni-Vincentelli 43]. We also believe that it is possible to optimize across model-of-computation boundaries to improve the performance of and reduce errors in the design at an early stage in the process. There are many di erent views on how to accomplish this. There are two essential approaches: one is to develop encapsulation techniques for each pair of models that allow di erent models of computation to interact in a meaningful way, i.e., data produced by one object are presented to the other in a consistent way so that the object \understands" 17]. The other is to develop an encompassing framework where all the models of importance \reside" so that their combination, re-partition and communication happens in the same generic framework. While we realize that today heterogeneous models of computation are a necessity, we believe that the second approach will be possible and will provide the designer with a powerful mechanism to actually select the appropriate models of computation, (e.g., FSMs, Data- ow, Discrete-Event, that then become a lower level of abstraction with respect to the uni ed model) for the essential parts of the design. 5

At this level is also important to orthogonalize concerns, that is to separate di erent aspects of the design. In this regard, a natural dividing line is the separation between functionality and communication. That is, we view a design as composed of functional behavior (modules) and communication behavior (between modules), which are themselves further decomposed as we re ne and analyze the design. It is our strong belief that communication is key in assembling systems on a chip from separate Hip's. Communication is a complex issue since the functionality of the components being interconnected must be preserved. The separation between function and communication will be emphasized throughout the paper. In addition, a model of computation that encompasses the key aspects of Discrete Event, Data-Flow and Finite State Machine models will be presented in detail. This model, called network of Co-design Finite State Machines (CFSM), is the backbone of the POLIS system developed at the University of California at Berkeley, an environment for function-architecture co-design with particular emphasis on control-dominated applications and on software development. (See Figure 2 for a block diagram of the functionalities of the environment.) This model is also used as the basic semantic model for an industrial product of Cadence Design Systems, an environment for embedded system design including multi-media and telecommunication applications. The paper is organized as follows. In Section 2, we present the mathematical machinery used to compare and describe the models of computation. In Section 3, we present and compare the most important models of computation. In Section 4, we introduce the CFSM model. In Section 5, we give some concluding remarks.

2 MOCs: Basic Concepts and the Tagged Signal Model

An MOC is composed of a description mechanism (syntax) and rules for computation of the behavior given the syntax (semantics). An MOC is chosen for describing a sub-behavior of a design based on its suitability: compactness of description, delity to design style, ability to synthesize and optimize the behavior to an appropriate implementation. For example, some MOCs are suitable for describing complicated data transfer functions and completely unsuitable for complex control, while others are designed with complex control in mind. There are a number of basic ideas and primitives that are commonly used in formulating models of computation. Most MOCs permit distributed system description (a collection of communicating modules), and give rules dictating how each module computes (function) and how they transfer information between them (communication). Some of the primitives include combination Boolean 6

2.1 Modeling Embedded Systems with MOCs

Formal Languages

Translators

System Behavior Co?Simulation Partitioning Architectural Selection Scheduler Selection Formal Verification Estimates

Scheduler Template + Timing Constraints

Partitioned Specification

Interface Synthesis

SW Synthesis

HW Synthesis

S?Graph

Unoptimized HW

HW Interfaces

Verif. Interm. Format

OS Synthesis

Task Synthesis

Logic Synthesis

HW Estimation SW Estimation

C?code

Optmized HW

Partitioning

Optmized HW

BOARD LEVEL PROTOTYPING

Standard Components

Physical Prototype

Figure 2: The POLIS design framework

7

functions and synchronous state machines for specifying function, and queues, bu ers, schedulers for specifying communication. Function and communication are often not described completely separately for e ciency and optimization. More precisely, MOCs are typically realized (implemented in practice) by a particular language and its semantics. We elaborate on the distinction between MOCs and languages in Section 2.2. A language is a set of symbols, rules for combining them (its syntax), and rules for interpreting combinations of symbols (its semantics). Two approaches to semantics have evolved, denotational and operational. A language can have both (ideally they are consistent with one another, although in practice this can be di cult to achieve). Denotational semantics, rst developed by Scott and Strachey 59], gives the meaning of the language in terms of relations. Operational semantics, which dates back to Turing machines, gives meaning of a language in terms of actions taken by some abstract machine, and is typically closer to the implementation. Models of computation can be viewed based on the following characteristics: the kinds of relations that are possible in a denotational semantics how the abstract machine behaves in an operational semantics how individual behavior is speci ed and composed how hierarchy abstracts this composition communication style A design (at all levels of the abstraction hierarchy from functional speci cation to nal implementation) is generally represented as a set of components, which can be considered as isolated monolithic blocks, which interact with each other and with an environment that is not part of the design. The model of computation de nes the behavior and interaction of these blocks. We view MOCs at two levels of abstraction. At the higher level, we take the view of the tagged signal model (which we call here TSM) described in section 2.3. The TSM abstraction de nes processes and their interaction using signals composed of partially ordered events, in turn composed of tags and values. We use processes to describe both functional behavior and communication behavior. This is a denotational view, though it is not associated with a particular language. We use this model to compare elements of di erent models of computation, styles of sequential behavior, concurrency, and communication at a high level. At the lower level of abstraction, we take the view of general primitives for function and timing (used in the re nement of TSM processes), where each MOC constitutes a particular choice of these two. This is a more operational view. We give precise de nitions for a number of terms, but these de nitions will inevitably con ict with standard usage in some communities. We have discovered 8

that, short of abandoning the use of most common terms, no terminology can be consistent with standard usage in all related communities. We attempt to avoid confusion by being precise, even at the risk of being pedantic. The basic primitive concepts are describe in Section 2.4. The primitive building blocks for speci cation and implementation are given in Section 2.5. All these basic primitives and concepts are then used in Section 3 to classify and describe the main MOCs that appear in the literature. The distinction between a language and its underlying model of computation is important. The same model of computation can give rise to fairly di erent languages (e.g., the imperative Algol-like languages C, C++, Pascal, and fortran). Some languages, such as VHDL and Verilog, support two or more models of computation1. The model of computation a ects the expressiveness of a language | which behaviors can be described in the language, whereas the syntax a ects compactness, modularity, and reusability. Thus, for example, object-oriented properties of imperative languages like C++ are more a matter of syntax than a model of computation. The expressiveness of a language is an important issue. A language that is not expressive enough to specify a particular behavior is clearly unsuitable, while a language that is too expressive is often too complex for analysis and synthesis. For very expressive languages, many analysis and synthesis problems become undecidable: no algorithm will solve all problem instances in nite time. A language in which a desired behavior cannot be represented succinctly is also problematic. The di culty of solving analysis and synthesis problems is at least linear in the size of the problem description, and can be as bad as several times exponential, so choosing a language in which the description of the desired behavior of the system is compact can be critical. A language may be very incomplete and/or very abstract. For example, it may specify only the interaction between computational modules, and not the computation performed by the modules. In this case, it provides an interface to a host language that speci es the computation, and is called a coordination language (examples include Linda 20], Granular Lucid 34], and Ptolemy domains 17]). Another language may specify only the causality constraints of the interactions without detailing the interactions themselves nor providing an interface to a host language. In this case, the language is used as a tool to prove properties of systems, as done, for example, in process calculi 33, 46] and Petri nets 50, 53]. In still more abstract modeling, components in the system are

1 They directly support the Imperative model within a process, and the Discrete Event model among processes. They can also support Extended Finite State Machines under suitable restrictions known as the \synthesizable subset".

2.2 Languages and Models of Computation

9

replaced with nondeterminate speci cations that give constraints on the behavior, but not the behavior itself. Such abstraction provides useful simpli cations that help formal veri cation. At the highest level of abstraction, we adopt the tagged-signal model (TSM) proposed by Lee and Sangiovanni-Vincentelli 42]. It is a formalism for describing aspects of models of computation for embedded system speci cation. It is denotational in the Scott and Strachey 59] sense, and it de nes a semantic framework (of signals and processes) within which models of computation can be studied and compared. It is very abstract|describing a particular model of computation involves imposing further constraints that make it more concrete. The fundamental entity in the TSM is an event: a value/tag pair. Tags are often used to denote temporal behavior. A set of events (an abstract aggregation) is a signal. Processes are relations on signals, expressed as sets of n-tuples of signals. A particular model of computation is distinguished by the order it imposes on tags and the character of processes in the model. More formally, given a set of values V and a set of tags T , an event is a member of T V . A signal s is a set of events, and thus is a subset of T V . A functional (or deterministic) signal is a (possibly partial) function from T to V . The set of all signals is denoted S . A tuple of n signals is denoted s, and the set of all such tuples is denoted S n . The di erent models of time that have been used to model embedded systems can be translated into di erent order relations on the set of tags T in the taggedsignal model. In a timed system T is totally ordered, i.e., there is a binary relation < on members of T such that if t1 ; t2 2 T and t1 6= t2 , then either t1 < t2 or t2 < t1. In an untimed system, T is only partially ordered.

2.3 The Tagged-Signal Model

2.3.1 Signals, tags and events

2.3.2 Processes

A process P with n signals is a subset of the set of all n-tuples of signals, S n for some n. A particular s 2 S n is said to satisfy the process if s 2 P . An s that satis es a process is called a behavior of the process (intuitively, it is the generalization of a \simulation trace"). Thus a process is a set of possible behaviors, or a constraint on the set of \legal" signals. Intuitively, processes in a system operate concurrently , and constraints imposed on their signal tags de ne communication 2 among them. The environment in which the system operates can be modeled with a process as well.

2 This is often called also synchronization, but we will try to avoid using the term in this sense because it is too overloaded.

10

For many (but not all) applications, it is natural to partition the signals associated with a process into inputs and outputs. Intuitively, the process does not determine the values of the inputs, and does determine the values of the outputs. If n = i + o, then (S i ; S o ) is a partition of S n . A process with i inputs and o outputs is a subset of S i S o . In other words, a process de nes a relation between input signals and output signals. A (i + o)-tuple s 2 S i+o is said to satisfy P if s 2 P . It can be written s = (s1 ; s2), where s1 2 S i is an i-tuple of input signals for process P and s2 2 S o is an o-tuple of output signals for process P . If the input signals are given by s1 2 S i , then the set I = f(s1; s2) j s2 2 S o g describes the inputs, and I \ P is the set of behaviors consistent with the input s1 . A process F is functional (or deterministic) with respect to an input/output partition if it is a single-valued, possibly partial, mapping from S i to S o . That is, if (s1; s2) 2 F and (s1; s3) 2 F , then s2 = s3. In this case, we can write s2 = F (s1), where F : S i ! S o is a (possibly partial) function. Given the input signals, the output signals are determined (or there is unambiguously no behavior). A process is completely speci ed if it is a total function, that is, for all inputs in the input space, there is a unique behavior. Process composition in the TSM is de ned by the intersection of the constraints each process imposes on each signal. To facilitate its de nition, we assume that all the processes that are composed are de ned on the same set of signals3. Hence a composition of a set of processes is also a process. In the rest of this paper, we will use processes to model both function and communication . Generally, an MOC de nes a exible mechanism for modeling function, and a rigid mechanism (signal, queue, shared variable, : : : ; see Section 2.5) for modeling communication. On the other hand, the TSM must compare di erent MOCs and hence be exible when modeling communication as well. It is, however, useful to distinguish between the two at least conceptually, since: 1. Functional processes are mostly concerned with the value component of their signals, and generally do not have much to do with the tag component. In other terms, the constraints a functional process imposes on its input and output signals are generally complex with respect to values, but much simpler with respect to tags. 2. Communication processes are solely concerned with the tag component of their signals, while values are left untouched.

3 This can be obtained trivially, since a process can be extended to any new signal by simply not imposing any constraint on it.

2.3.3 Process composition

11

One of the most useful and important questions to ask when composing processes, is what properties of the isolated processes are preserved by composition . Here we focus only on two fundamental properties: functionality (unique output n-tuple for every input n-tuple) and complete speci cation (for every input n-tuple there exists a unique output n-tuple). To analyze this aspect, we note that, given a formal model of the functional speci cations and of the properties, three situations may arise: 1. The property is inherent for the model of the speci cation (i.e., it can be shown formally to hold for all speci cations described using that model). 2. The property can be veri ed syntactically for a given speci cation (i.e., it can be shown to hold with a simple, usually polynomial-time, analysis of the speci cation). 3. The property must be veri ed semantically for a given speci cation (i.e., it can be shown to hold by executing, at least implicitly, the speci cation for all inputs that can occur). For example, consider the functionality property. Any design described by a data ow network (a formal model to be described later) is functional (also called deterministic or determinate in data- ow vernacular), and hence this property need not be checked for this model of computation. If the design is represented by a network of FSMs (for example, synchronous composition of Mealy Finite State Machines), even if the components are functional and completely speci ed, the result of the composition may be either incompletely speci ed (the composition has no solution) or non-functional (the composition has multiple solutions). These situations arise if and only if when a combinational feedback loop exists in the composition: with an odd number of Boolean inverters, there is no \solution" and the composition is incompletely speci ed, with an even number of inverters, there are multiple solutions and the composition is non-functional. A syntactical check on the composition to verify whether combinational loops exist can be carried out. If none exist, then the composition is functional and completely speci ed. With Petri nets, on the other hand, functionality is di cult to prove: it must be checked by exhaustive simulation. Consider, as a motivating example introducing these several mechanisms to denote temporal behavior, two problems: one of analysis, modeling a timeinvariant dynamical system on a computer, and one of design, the design of a two-elevator system controller.

2.3.4 Examples

Analysis Example The underlying mathematical model of a time-invariant

dynamical system, a set of di erential equations over continuous time, is not 12

directly implementable on a digital computer, due to the double quantization of real numbers into nite bit strings, and of time into clock cycles. Hence a rst translation is required, by means of an integration rule , from the di erential equations to a set of di erence equations , that are used to compute the values of each signal with a given tag from the values of some other signals with previous and/or current tags. If it is possible to identify several strongly connected components in the dependency graph4, then the system is decoupled . It becomes then possible to go from the total order of tags implicit in physical time to a partial order imposed by the depth- rst ordering of the components. This partial ordering gives us some freedom in implementing the integration rule on a computer. We could, for example, play with scheduling by embedding the partial order into the total order among clock cycles. It is often convenient, for example, to evaluate a component completely, for all tags, before evaluating components that depend on it. It is also possible to spread the computation among multiple processors. In the end, time comes back into the picture, but the double transformation , from total to partial order, and back to total order again, is essential to 1. prove properties about the implementation (such as stability of the integration method, or a bound on the maximum execution time), 2. optimize the implementation with respect to a given cost function (e.g., size of the bu ers required to hold intermediate signals versus execution time, or satisfaction of a constraint on the maximum execution time). was to avoid over-speci cation of designs. For a two-elevator system controller, a simplistic set of speci cations can be expressed as follows: respond to all requests in the exact order they are received with the criterion that the maximum delay from the time a request is received and the time the elevator service is o ered, is minimized. It is clear that the two elevators are concurrent subsystems and that their operation can be controlled with no need to \synchronize" their operation. It is determined by analyzing the order of events not the exact time of occurrence. However, if no assumption is made about the way requests are made, then we may end up in a dead-lock situation due to the nature of the speci cation. In fact, if three requests are made at the same instant of time, then the response cannot follow the speci cation, since there are only two elevators available. One solution to this problem is to assume that no two requests may happen at the exact same time; then the speci cation can be met for the two elevator system. Another solution is to arbitrarily assign priorities among requests that happen at the same time: the speci cation is changed to re ect these priorities instead of those implied by the order of occurrence. The most

4 A directed graph with a node for each signal, and an edge between two signals whenever the equation for the latter depends on the former.

Design Example. One of the key motivations for the tagged signal model

13

important aspect in design is to capture the intent of the designer by abstracting away the non-essential aspects of the system. This example illustrates that it is essential to classify MOCs by their treatment of events with the same tag. This aspect is strictly related to the notion of synchrony and asynchrony as we will see later. Once the control algorithm has been developed, then its implementation needs to be carried out. If the algorithm is implemented in software running on a single processor, then all events processed are totally ordered and the order is determined by the intrinsic order coming from the speci cations (order of occurrence of the requests) and by the existence of limited resources. Even if the partial order dictated by the algorithm exposes some potential parallelism, the presence of a single processor forces sequential execution determined by a scheduling algorithm that decides in which order the operations are executed. Hence, in the end, we need to map an abstract design into the physical world characterized by real time and limited resources that imposes a global ordering on events.

2.4 Comparing Models of Computation

A TSM process is, according to the de nition given, a partial mapping from input signals to output signals. In order to consider more concrete mappings, we introduce some primitive concepts on which they are based. System behavior, as we have previously stated, is composed of functional behavior and communication behavior, each represented by TSM processes. A process in turn is composed of functional behavior and timing behavior. Function is how things happen, or in the TSM, how events are related (how inputs are used to compute outputs) \around" a particular tag. Time is the order in which things happen, or in the TSM, the assignment of a tag to each event. The distinction between function and time is not this clean in every context. For example, a state in a nite state machine cannot be labeled as belonging exclusively to the function or time component of the behavior of the machine, but is rather based on the history of both. Nonetheless, the division between function and time, particularly at a primitive level, is useful in the conception and understanding of MOCs. System operation can be viewed as a series of process computations, sometimes called rings. We will use function, time, computation ( ring) to describe MOCs and their primitives. In the sections that follow, we consider the fundamental concepts which are used to re ne our processes. For functional processes, we will rst consider state-less processes in which only inputs with a given tag concur to form outputs with the same tag. We then introduce the notion of state in the context of process networks. For communication processes, we then consider the primitives for concurrency and inter-process communication. Finally, we give the basic building blocks used to realize these concepts in practice. It is from these that 14

today's most prevalent models of computation are built.

2.4.1 Process function

In the control-dominated arena, since the pioneering work of Shannon, Boolean functions have been used as a representation of both a system speci cation and its implementation in hardware (relay networks in Shannon's time, CMOS gates now). Several formally equivalent (but often with di erent levels of convenience in practice) representations for binary- and multi-valued boolean functions have been proposed, such as: truth table, Boolean network 57], which is a Directed Acyclic Graph (DAG), with a truth table associated with each node and edges carrying Boolean variable values, Binary Decision Diagram 15], that is also a DAG with one level of nodes for each input variable, and each node acting as a \multiplexer" between the function values associated with every variable values. In the data-dominated arena, Data Flow actors play the role of processes and represent functions from simple state-less arithmetic operations such as addition and multiplication, to higher-level \combinational" transformations, such as Fast Fourier Transform. Most models of computation include components with state, where behavior is given as a sequence of state transitions. State in a process network can always be simply implemented by means of feedback. An output and an input signal can be connected together, and thus provide a connection between process inputs and outputs beyond the tag barrier. However, we can also consider a notion of state within a process, since this can be useful in order to \hide" the implementation of the state information. We can formalize this notion by considering a process F that is functional with respect to partition (S i ; S o ). Let us assume for the moment that F belongs to a timed system, in which tags are totally ordered5 . Then for any tuple of signals s, we can de ne s>t to be a tuple of the (possibly empty) subset of the events in s with tags greater than t. Two input signal tuples r; s 2 S i are in relation EtF (denoted (ri ; si ) 2 EtF ) if r>t = s>t implies F (r)>t = F (s)>t. This de nition intuitively means that process F cannot distinguish between the \histories" of r and s prior to time t. Thus, if the inputs are identical after time t, then the outputs will also be identical.

5

2.4.2 Process State

A de nition of state for untimed systems is also possible, but it is much more involved.

15

EtF is obviously an equivalence relation, partitioning the set of input signal tuples into equivalence classes for each t. Following a long tradition, we call these equivalence classes the states of F . In the hardware community, components with only one state for each t are called combinational , while components with more than one state for some t are called sequential . Note however that the term \sequential" is used in very di erent ways in other communities.

The sequential or combinational behavior just described is related to individual processes, and embedded systems will typically contain several coordinated concurrent processes. At the very least, such systems interact with an environment that evolves independently, at its own speed. It is also common to partition the overall model into tasks that also evolve more or less independently, occasionally (or frequently) interacting with one another. This interaction implies a need for coordinated communication. Communication between processes can be explicit or implicit. Explicit communication implies forcing an order on the events, and this is typically realized by designating a sender process which informs one or more receiver processes about some part of its state. Implicit communication implies the sharing of tags (i.e., of a common time scale), which forces a common partial order of events, and a common notion of state. The problem with this form of communication is that it must be physically implemented via shared signals (e.g., a common reference clock), whose distribution may be di cult in practice.

2.4.3 Concurrency and Communication

Basic Time Time plays a larger role in embedded systems than in classical computation. In classical transformational systems, the correct result is the primary concern|when it arrives is less important (although whether it arrives, the termination question, is important). By contrast, embedded systems are usually real-time systems, where the time at which a computation takes place is very important. As mentioned previously, di erent models of time become di erent order relations on the set of tags T in the tagged-signal model. Recall that in a timed system T is totally ordered, while in an untimed system T is only partially ordered. Implicit communication generally requires totally ordered tags, usually identi ed with physical time. The tags in a metric-time system have the notion of a \distance" between them, much like physical time. Formally, there exists a partial function d : T T ! R mapping pairs of tags to real numbers such that d(t1 ; t2) = 0 , t1 = t2, d(t1; t2) = d(t2 ; t1) and d(t1; t2) + d(t2; t3) >= d(t1; t3). Two events are synchronous if they have the same tag (the distance between them is 0). Two signals are synchronous if each event in one signal is synchronous with an event in the other signal and vice versa.

16

Treatment of Time in Systems A discrete-event system is a timed system

where the tags in each signal are order-isomorphic with the natural numbers 42]. Intuitively, this means that any pair of ordered tags has a nite number of intervening tags. This is the basis of the underlying MOC of the Verilog and VHDL hardware description languages 62, 49]. A synchronous system is one in which every signal in the system is synchronous with every other signal in the system. A discrete-time system is a synchronous discrete-event system. An asynchronous system is a system in which no two events can have the same tag. If tags are totally ordered, the system is asynchronous interleaved , while if tags are partially ordered, the system is asynchronous concurrent . For asynchronous systems concurrency and interleaving are, to a large extent, interchangeable, since interleaving can be obtained from concurrency by embedding the partial order into a total order, and concurrency can be reconstructed from interleaving by identifying \untimed causality" 48]. Note that time is a continuous quantity. Hence real systems are asynchronous by nature. Synchronicity is only a (very) convenient abstraction, that may be expensive to implement due to the need to share tags, and hence, as discussed above, to share a reference \clock" signal. Synchronous/reactive languages (see e.g. 28]) deserve special mention. They have an underlying synchronous model in which the set of tags in any behavior of the system implies a global \clock" for the system. However, to make this MOC synchronous in the sense of the TSM, we need to assume that every signal conceptually has an event at every tag. In some synchronous/reactive designs this may not be the case but if we de ne the events in the process to include a value denoting the absence of an event, then all synchronous/reactive models can be de ned as synchronous in our framework. At each clock tick, each process maps input values to output values. Note that if we include the absent value for events, then discrete-event systems are also synchronous. The main di erences are: in the granularity of tags: intuitively, synchronous models should be used for systems in which there are fewer tags, and in the number of events that have the absent value at any tag: intuitively, synchronous models should be used for systems in which only few events have absent values. Particular attention has to be devoted to events with values at the same tag and that have cyclic dependencies ("combinational cycles"). The existence of such dependencies implies that the input-output relation is described implicitly as the solution of an algebraic set of equations. This set of equations may have either a single solution for each input value, in which case the process is completely speci ed, no solution for some input value, in which case the process is functional but not completely speci ed, or multiple solutions for some 17

input value, in which case the process is not functional. This is the source of endless problems in systems described by VHDL or Verilog and di culties in synchronous/reactive languages. A possibility when facing a cyclic dependency is to leave the result unspeci ed, resulting in nondeterminacy or, worse, in nite computation within one tick according to the particular input values (VHDL, Verilog and some variants of StateCharts belong to this class 65]). A better approach is to use xed-point semantics, where the behavior of the system is de ned as a set of events that satisfy all processes 11]. Given this approach to the problem, there are procedures that can determine the existence of single or multiple xed points in nite time, thus avoiding nasty inconsistencies and di culties. Asynchronous systems do not su er from this problem since there cannot be cyclic dependencies at the same tag given that only one event can have a value at any given tag. Note that often asynchronous systems are confused with discreteevent systems and thus it is not infrequent to nd assertions in the literature that asynchronous systems may have inconsistent or multiple solutions when indeed this is never the case!

Implementation of Concurrency and Communication Concurrency in

physical implementations of systems implies a combination of parallelism, which employs physically distinct computational resources, and interleaving, which means sharing of a common physical resource. Mechanisms for achieving interleaving, generally called schedulers , vary widely, ranging from operating systems that manage context switches to fully-static interleaving in which multiple concurrent processes are converted (compiled) into a single process. We focus here on the mechanisms used to manage communication between concurrent processes. Parallel physical systems naturally share a common notion of time, according to the laws of physics. The time at which an event in one subsystem occurs has a natural ordering relationship with the time at which an event occurs in another subsystem. Physically interleaved systems also share a natural common notion of time: one event happens before another and the time between them can be computed (of course, accuracy is an issue). Logical systems, on the other hand, need a mechanism to explicitly share a notion of time (communicate ). Consider two imperative programs interleaved on a single processor under the control of a time-sharing operating system. Interleaving creates a natural ordering between events in the two processes, but this ordering is generally unreliable, because it heavily depends on scheduling policy, system load and so on. Some explicit communication mechanism is required for the two programs to cooperate. One way of implementing this could be by forcing both to operate based on a global notion of time, which in turn forces a total order on events. This can be extremely expensive. In practice, this communication is done explicitly, where the total order is replaced by a partial 18

order. Returning to the example of two processes running under a time-sharing operating system, we take precautions to ensure an ordering of two events only if the ordering of these two events matters. We can do this by communicating through common signals, and forcing one process to wait for a signal from the other, which forces the scheduler to interleave the processes in a particular way. A variety of mechanisms for managing the order of events, and hence for communicating information between processes, exists. We will now examine and classify them according to the tagged-signal model, by using \special-purpose" processes to model communication. Using processes to model communication (rather than considering it as \primitives" of the tagged-signal model) makes it easier to compare di erent MOCs, and also allows one to consider re ning these communication processes when going from speci cation to implementation 56]. Recall that the communication primitive in the TSM is the event, which is a two-component entity whose value is related to function and whose tag is related to time. That is, communication is implemented by two operations: 1. the transfer of values between processes (function; TSM event value), 2. the determination of the relationship in time between two processes (time; TSM event tag).

Unfortunately, often the term \communication" (or data transfer) is used for the former, and the term \synchronization" is used for the latter. We feel, however, that the two are intrinsically connected in embedded systems: both tag and value carry information about a communication. Thus, communication and synchronization, as mentioned before, are terms which cannot really be distinguished in this sense.

2.5 Basic communication primitives

In this section, we de ne some of the communication primitives that have been described in the literature, following the classi cation developed in the previous sections. Unsynchronized In an unsynchronized communication, a producer of information and a consumer of the information are not coordinated. There is some connection between them (e.g., a bu er) but there is no guarantee that the consumer reads \valid" information produced by the producer, and no guarantee that the producer will not overwrite previously produced data before the consumer reads the data. In the tagged-signal model, the repository for the data is modeled as a process, and the reading and writing actions are modeled as events without any enforced ordering of their tags. Read-modify-write Commonly used for accessing shared data structures in software, this strategy locks a data structure during a data access (read, 19

write, or read-modify-write), preventing any other accesses. In other words, the actions of reading, modifying, and writing are atomic (indivisible, and thus uninterruptible). In the tagged-signal model, the repository for the data is modeled as a process where events associated with this process are totally ordered (resulting in a partially ordered model at the global level). The read-modify-write action is modeled as a single event. Unbounded FIFO bu ered This is a point-to-point communication strategy, where a producer generates (writes) a sequence of data tokens and a consumer consumes (reads) these tokens, but only after they have been generated (i.e., only if they are valid). In the tagged-signal model, this is a simple connection where the signal on the connection is constrained to have totally ordered tags. The tags model the ordering imposed by the FIFO model. If the consumer process has unbounded FIFOs on all inputs, then all inputs have a total order imposed upon them by this communication choice. This model captures essential properties of both Kahn process networks and data ow 35]. Bounded FIFO bu ered In this case (we discuss only the point-to-point case for the sake of simplicity), the data repository is modeled as a process that imposes ordering constraints on its inputs (which come from the producer) and the outputs (which go to the consumer). Each of the input and output signals are internally totally ordered, while their combination is partially ordered. The simplest case is where the size of the bu er is one, in which case the input and output events must be perfectly interleaved (i.e., that each output event lies between two input events). Larger bu ers impose a maximum di erence (often called synchronic distance 51]) between the number of input or output events occurring in succession. Note that some implementations of this communication mechanism may not really block the writing process when the bu er is full, thus requiring some higher level of ow control to ensure that this never happens, or that it does not cause any harm. Petri net places This is a multi-partner communication strategy, where several producers generate tokens and several consumers consume these tokens 51]. In the tagged-signal model, this is modeled as a process that keeps track of the tags of its input (from producers) and output (to consumers) signals. As in the previous case, each signal has totally ordered events, and the process makes sure that the number of input events is always greater than or equal to that of output events. Rendezvous In the simplest form of rendezvous, which is embodied in the underlying MOC of the Occam and Lotos 64] languages, a single writing process and a single reading process must simultaneously be at the point in their control ow where the write and the read occur. It is a 20

Transmitters Receivers Unsynchronized Read-Modify-Write Unbounded FIFO Bounded FIFO Petri net place Single Rendezvous Multiple Rendezvous many many one one many one many many many one one many one many

Bu er Blocking Blocking Single Size Reads Writes Reads one no no no one yes yes no unbounded yes no yes bounded maybe maybe yes unbounded no no yes one yes yes yes one no no yes

Table 1: A comparison of concurrency and communication schemes. convenient communication mechanism, because it has the semantics of a single assignment, in which the writer provides the right-hand side, and the reader provides the left-hand side. In the tagged-signal model, this is imposed by events with identical tags 42]. Lotos o ers, in addition, multiple rendezvous, in which one among multiple possible communications is non-deterministically selected. Multiple rendezvous is more exible than single rendezvous, because it allows the designer to specify more easily several \expected" communication ports at any given time, but it is very di cult and expensive to implement correctly. Of course, various combinations of the above models are possible. For example, in a model that partially uses the unsynchronized communication scheme, a consumer of data may be required to wait until the rst time a producer produces data, after which the communication is unsynchronized. The essential features of the concurrency and communicationstyles described above are presented in Table 1. These are distinguished by the number of transmitters and receivers (e.g., broadcast versus point-to-point communication), the size of the communication bu er, whether the transmitting or receiving process may continue after an unsuccessful communication attempt (blocking reads and writes), and whether the result of each write can be read at most once (single reads). Note that, strictly speaking, the blocking/nonblocking read and write aspects are part of the \functional" processes, and not of the \communication" processes. However, these communication schemes also specify that aspect, and hence we chose to include in the table. A \maybe" entry means that MOCs considering both the \yes" and \no" answer have been proposed in the literature.

3 Common Models of Computation

We are now ready to use the scheme developed in the previous Section to classify and analyze several models of computation that have been used to describe 21

embedded systems. We will consider issues such as ease of modeling, e ciency of analysis (simulation or formal veri cation), automated synthesizability, and optimization space versus over-speci cation. We assume a background knowledge of basic, non-concurrent MOCs such as Finite Automata, Turing Machines, and Algebraic State Machines, and we focus on the timing, concurrency and communication aspects instead.

3.1 Discrete-Event

Time is an integral part of a discrete-event model of computation. Events usually carry a totally-ordered time stamp indicating the time at which the event occurs. A DE simulator usually maintains a global event queue that sorts events by time stamp. Digital hardware is often simulated using a discrete-event approach. The Verilog language 62], for example, was designed as an input language for a discrete-event simulator. The VHDL language 49] also has an underlying discrete-event model of computation. Discrete-event modeling can be expensive|sorting time stamps can be timeconsuming. Moreover, ironically, although discrete-event is ideally suited to modeling distributed systems, it is very challenging to build a distributed discreteevent simulator. The global ordering of events requires tight coordination between parts of the simulation, rendering distributed execution di cult. Discrete-event simulation is most e cient for large systems with large, frequently idle or autonomously operating sections. Under discrete-event simulation, only the changes in the system need to be processed, rather than the whole system. As the activity of a system increases, the discrete-event paradigm becomes less e cient because of the overhead inherent in processing time stamps. Simultaneous events, especially those arising from zero-delay feedback loops, present a challenge for discrete-event models of computation. In such a situation, events may need to be ordered, but are not. Consider the discrete-event system shown in Figure 3. Process B has zero delay, meaning that its output has the same time stamp as its input. If process A produces events with the same time stamp on each output, there is ambiguity about whether B or C should be invoked rst, as shown in Figure 3(a). Suppose B is invoked rst, as shown in Figure 3(b). Now, depending on the simulator, C might be invoked once, observing both input events in one invocation, or it might be invoked twice, processing the events one at a time. In the latter case, there is no clear way to determine which event should be processed rst. The problem could be solved by requiring the user to provide a delay for each process, but this is not convenient in general. Hence various simulators have resorted to various heuristic techniques: The VHDL simulation semantics 49] uses a synchronous model (with unit 22

t

A

t

C A B (b) C

t

B (a)

t

C

t t+

A

A

B (c)

B (d)

t+

C

Figure 3: Simultaneous events in a discrete-event system. (a) Process A produces events with the same time stamp. Should B or C be red next? (b) Zero-delay process B has red. How many times should C be red? (c) Deltadelay process B has red; C will consume A's output next. (d) C has red once; it will re again to consume B's output. delay, called \delta step") in order to provide a two-level structure of time and thus solve non-determinism within a given \real time" instant. Each instant of time (level 1) is broken into (a potentially in nite number of) totally ordered delta steps (level 2). A \zero-delay" process in this model actually has delta steps, or ordered progress towards a solution though no real time elapses. For example, if Process B contains a delta step between input and output, ring A followed by B would result in the situation in Figure 3(c). The next ring of C will see the event from A only; the ring after that will see the (delay-ordered) event from B. The Discrete Event domain in Ptolemy 17] uses a synchronous model, but with mostly zero delay and only enough delta steps to eliminate all zero-delay cycles. The BONES simulator by Cadence uses an asynchronous model. Adding a feedback loop from Process C to A in Figure 3 would create a problem if events circulate through the loop without any increment in time stamp. The same problem occurs in synchronous languages, where such loops are called causality loops. No precedence analysis can resolve the ambiguity. In synchronous languages, the compiler may simply fail to compile such a program. 23

Discrete-event simulators attempt to identify such cases and report them to the user. We wish to stress that delta steps do not have a meaning of time (though they are often called delta \delay"). They are just a clever mechanism to implement a xed point computation used to compute the behavior of the system at a point in time. Fixed point iteration can also be used in the synchronous/reactive model to de ne its semantics and make it determinate. Hence \delta steps" can also be thought of as an \iteration index". Moreover, VHDL uses an event model that is not monotonic, and hence the xed point may never be reached, as discussed above. On the other hand, synchronous language use a ternary logic model, in which xed point convergence in ensured in a nite number of steps 16]. The reason why DE is a popular MOC in practice is that it has been implemented e ciently in a number of event-driven simulators, and it is quite convenient to evaluate the performance of very large and complex systems. By imposing little restriction on the modeling style, it makes simulation simple and synthesis as well as formal veri cation hard. In data ow, a program is speci ed by a directed graph where the nodes (called actors) represent computations and the arcs represent totally ordered sequences (called streams) of events (called tokens). In gure 4(a), the large circles represent actors, the small circle represents a token and the lines represent streams. The graphs are often represented visually and are typically hierarchical, in that a node in a graph may represent another directed graph. The nodes in the graph can be either language primitives or subprograms speci ed in another language, such as C or fortran. In the latter case, we are actually mixing two models of computation, where data ow serves as the coordination language for subprograms written in an imperative host language. Data ow is a special case of Kahn process networks 35, 41]. In a Kahn process network, communicationis by unbounded FIFO bu ering, and processes are constrained to be continuous mappings from input streams to output streams. \Continuous" in this usage is a topological property that ensures that the program is determinate 35]. Intuitively, it implies a form of causality without time; speci cally, a process can use partial information about its input streams to produce partial information about its output streams. Adding more tokens to the input stream will never result in having to change or remove tokens on the output stream that have already been produced. One way to ensure continuity is with blocking reads, where any access to an input stream results in suspension of the process if there are no tokens. One consequence of blocking reads is that a process cannot test an input channel for the availability of data and then branch conditionally to a point where it will read a di erent input. In data ow, each process is decomposed into a sequence of rings, indivisible 24

3.2 Data ow Process Networks

quanta of computation. Each ring consumes and produces tokens. Dividing processes into rings avoids the multi-tasking overhead of context switching in direct implementations of Kahn process networks. In fact, in many of the signal processing environments, a major objective is to statically (at compile time) schedule the actor rings, achieving an interleaved implementation of the concurrent model of computation. The rings are organized into a list (for one processor) or set of lists (for multiple processors). Figure 4(a) shows a data ow graph, and Figure 4(b) shows a single processor schedule for it. This schedule is a list of rings that can be repeated inde nitely. One cycle through the schedule should return the graph to its original state (here, state is de ned as the number of tokens on each arc). This is not always possible, but when it is, considerable simpli cation results 12]. In many existing environments, what happens within a ring can only be speci ed in a host language with imperative semantics, such as C or C++. A B A B (b) Figure 4: (a) A data ow process network (b) A single-processor static schedule for it A useful formal device is to constrain the operation of a ring to be functional, i.e., a simple, stateless mapping from input values to output values. Note, however, that this does not constrain the process to be stateless, since it can maintain state in a self-loop: an output that is connected back to one of its inputs. An initial token on this self-loop provides the initial value for the state. Many possibilities have been explored for precise semantics of data ow coordination languages, including Karp and Miller's computation graphs 37], Lee and Messerschmitt's synchronous data ow graphs 40], Lauwereins et al.'s cyclo-static data ow model 39, 13], Kaplan et al.'s Processing Graph Method (PGM) 36], Granular Lucid 34], and others 1, 20, 22, 60]. Many of these limit expressiveness in exchange for formal properties (e.g., provable liveness and bounded memory). 25 C D

(a) C D

Synchronous data ow (SDF) and cyclo-static data ow require processes to consume and produce a xed number of tokens for each ring. Both have the useful property that a nite static schedule can always be found that will return the graph to its original state. This admits extremely e cient implementations 12]. For more general data ow models, it is undecidable whether such a schedule exists 18]. A looser model of data ow is the tagged-token model, in which the partial order of tokens is explicitly carried with the tokens 3]. A signi cant advantage of this model is that while it logically preserves the FIFO semantics of the channels, it permits out-of-order execution. Some examples of graphical data ow programming environments intended for signal processing (including image processing) are Khoros 52], and Ptolemy 17]. Petri nets 50, 53] are, in their basic form, an in nite state model (just like data ow) for which, however, most properties are decidable in nite time and memory. They are interesting as an uninterpreted model for several very different classes of problems, including some relevant to embedded system design (e.g., process control, asynchronous communication, and scheduling). Moreover, a large user community has developed an impressive body of theoretical results and practical design aids and methods based on Petri nets. In particular, partial order-based veri cation methods (e.g. 63], 27], 45]) are one possible answer to the state explosion problem plaguing Finite State Machinebased veri cation techniques. A Petri net (PN) is a directed bipartite graph N = fP; T; F g. P is a set of places holding the distributed state (via tokens) of the system. T is a set of transitions, denoting the activity of the system. F P T T P is the ow relation, from places to transitions and vice-versa. Nodes linked by F are said to be in a predecessor/successor relationship. Transitions are often labeled with statements in a host language, just as in the case of DF actors. The state of the PN is the marking of the places, that is a non-negative integer valuation (\token assignment") of each place. The dynamic evolution of the PN is determined by the ring process of transitions. A transition may re whenever all its predecessor places are marked, and if it res, it decrements the marking (removes a token) of each predecessor and increments the marking of each successor (adds a token). PNs are interesting in general, and in particular in embedded system design, because they are a very general model of control, potentially with in nite state, yet very powerful analysis techniques, both exact and approximate, have been de ned for them. In particular, the ring rule of a PN bears a strong connection with linear algebra. If we represent the graph of the ow relation (given arbitrary orderings 26

3.3 Petri nets

p0 t1

t2

p1

p2

p3

p4

t3

t4

t5

t6

p1 p6

t3 t4 p2 p5

p5 p6 t5

t6

p4 p5 t5

t4 p5 p6

p3 p6 t0

t3 t0 p1 p2 t1 p0

t6 t2

p3 p4

(a)

(b)

Figure 5: Example of Free Choice Petri net and its Reachability Graph of the sets T and P ) as an incidence matrix I , and if we represent the current marking as an integer vector M , we can model the e ect of a sequence of transitions starting from M as follows. Let us denote by f the \ ring vector" of , that is a vector whose i-th position contains the number of times the i-th transition appears in . The marking M 0 reached after is given by

M 0 = If + M For example, consider the PN in Figure 5.(a), whose set of reachable markings is shown in Figure 5.(b). Its incidence matrix (one row for each place and one column for each transition) is: 1 ?1 ?1 0 0 0 0 0 1 0 ?1 0 0 0 0 1 0 0 ?1 0 0 0 0 1 0 0 ?1 0 0 0 1 0 0 0 ?1 ?1 0 0 1 0 1 0 ?1 0 0 0 1 0 1 The rst line corresponds to place p0, and has a 1 in position 0, because t0 adds 1 token to place p0, and ?1 in positions 1 and 2 because t1 and t2 remove one token from it. Consider now ring sequence = t0; t1; t3 whose ring vector (transposed) is f = j1101000jt. The marking M 0 reached from the initial marking M =

27

0 0 0 0 1 0 0 0 1 0 ?1 0 0 0 1 0 1 0 1 0 0 ?1 0 0 0 0 0 = 0 0 1 0 0 ?1 0 1 + 0 0 0 0 1 0 0 0 ?1 0 0 1 ?1 0 0 1 0 1 0 0 1 0 ?1 0 0 0 1 0 1 0 1 that corresponds to p2; p5 being marked, as expected. As another example, consider the initial marking with two tokens each in p5; p6 and the t ring sequence = t0; t0; t1; t3. In that case, M = j0000022jt, f = j2110000j t, and M 0 = j0111100j . This equation provides an interesting characterization of sequences of transitions that, when red from a marking M , return the net to same M . These sequences, also called T-invariants , must be solutions to 0 = If This is only a necessary condition, of course, since the sequences must also be reable from M (some intermediate step may yield a negative marking), but it is useful, e.g., when proving liveness conditions (e.g., showing that some transition can re in nitely often) or schedulability properties 47]. For example, in Figure 5 ring sequence t0; t1; t3; t4is a T-invariant j1101100jt that happens to be reable from the initial marking. The reader can check that this invariant is indeed a solution of the equation shown above. By duality (a very useful concept in Petri nets, based on exchanging the roles of places and transitions), one can also identify sets of places whose total cumulative marking cannot be changed by any ring sequence of the net. These sets, also called S-invariants can be used to establish the unreachability of a given marking, if it cannot be expressed as a linear combination of a basis of S-invariants 25]. Hence they can be very useful in proving (but not disproving) safety properties (e.g., the fact that some \dangerous" marking cannot be reached). Invariant-based techniques become necessary and su cient for a restricted (but expressive) class of PNs called free-choice nets 24], in which a multisuccessor place must be the only predecessor of its successors. The net of Figure 5.(a) is a free-choice net, since the only multi-successor place (p0) has only single-predecessor successors (t1; t2). In addition, reachability-based techniques for analysis, based on building the complete state space (or deciding in nite time that it is actually in nite), can also be used to prove properties of a given PN. 28

j0000011jt after ring is: 0 1 ?1 ?1

The basic PN model is interesting but somewhat limited in expressive power6 . For this reasons, people have extended it in various ways, such as adding colors to tokens. Colored PNs are similar to Data ow networks (with places playing the role of FIFOs and transitions playing the role of actors), but allow multiple predecessors and successors for a place/FIFO. In this way, they lose one of the most interesting properties of DF networks, determinacy , and of course gain something in terms of compactness and expressiveness7 . Time can also play an explicit role in PNs. Time has been associated with transitions and places, in various combinations and forms ( 55]). Generally speaking, time is associated with tokens, that carry a time stamp, and time stamps determine when transitions may re (and thus create new tokens with new time stamps). The problem with timed PNs is, as usual with real-time MOCs, that they su er from a particularly serious combinatorialexplosion problem when reducing the originally in nite timed state space to a nite set of equivalence classes, as discussed more in detail in Section 3.8. In a synchronous model of computation, all events are synchronous, i.e., all signals have events with identical tags. The tags are totally ordered, and globally available. Unlike the discrete-event model, all signals have events at all clock ticks, simplifying the simulator by requiring no sorting. Simulators that exploit this simpli cation are called cycle-based or cycle-driven simulators. Processing all events at a given clock tick constitutes a cycle. Within a cycle, the order in which events are processed may be determined by data precedences, which de ne the delta steps. These precedences are not allowed to be cyclic, and typically impose a partial order (leaving some arbitrary ordering decisions to the scheduler). Cycle-based models are excellent for clocked synchronous circuits, and have also been applied successfully at the system level in certain signal processing applications. A cycle-based model is ine cient for modeling systems where events do not occur at the same rate in all signals. While conceptually such systems can be modeled (using, for example, special tokens to indicate the absence of an event), the cost of processing such tokens is considerable. Fortunately, the cycle-based model is easily generalized to multirate systems. In this case, every nth event in one signal aligns with the events in another. A multirate cycle-based model is still somewhat limited. It is an excellent model for synchronous signal processing systems where sample rates are related by constant rational multiples, but in situations where the alignment of events in di erent signals is irregular, it can be ine cient.

6 It is more powerful than regular grammars, is incomparable with context-free grammars, and is less powerful than Turing machines. 7 The formal power is the same, being that of Turing machines for both general CPN and general DF.

3.4 Synchronous/Reactive

29

The more general synchronous/reactive model is embodied in the so-called synchronous languages 8]. Esterel 14] is a textual imperative language with sequential and concurrent statements that describe hierarchically-arranged processes. Lustre 29] is a textual declarative language with a data ow avor and a mechanism for multirate clocking. Signal 9] is a textual relational language, also with a data ow avor and a more powerful clocking system. Argos 44], a derivative of Harel's Statecharts 30], is a graphical language for describing hierarchical nite state machines (described more in detail in the next section). Halbwachs 28] gives a good summary of this group of languages. The synchronous/reactive languages describe systems as a set of concurrentlyexecuting synchronized modules. These modules communicate through signals that are either present or absent in each clock tick. The presence of a signal is called an event, and often carries a value, such as an integer. Most of these languages are static in the sense that they cannot request additional storage nor create additional processes while running. This makes them well-suited for bounded and speed-critical embedded applications, since their behavior can be extensively analyzed at compile time. This static property makes a synchronous program nite-state, greatly facilitating formal veri cation. Verifying that a synchronous program is causal (non-contradictory and deterministic) is a fundamental challenge with these languages. Since computation in these languages is delay-free and arbitrary interconnection of processes is possible, it is possible to specify a program that has either no interpretation (a contradiction where there is no consistent value for some signal) or multiple interpretations (some signal has more than one consistent value). Both situations are undesirable, and usually indicate a design error. A conservative approach that checks for causality problems structurally ags an unacceptably large number of programs as incorrect because most will manifest themselves only in unreachable program states. The alternative, to check for a causality problem in any reachable state, can be expensive since it requires an exhaustive check of the state space of the program. In addition to the ability to translate these languages into nite-state descriptions, it is possible to compile these languages directly into hardware. Techniques for translating both Esterel 10] and Lustre 54] into hardware have been proposed. The result is a logic network consisting of gates and ip- ops that can be optimized using traditional logic synthesis tools. To execute such a system in software, the resulting network is simply simulated. The technique is also the basis to perform more e ciently causality checks, by means of implicit state space traversal techniques 58]. Finite State Machines (FSMs) are an attractive model for embedded systems. The amount of memory required by such a model is always decidable, and is often an explicit part of its speci cation. Halting and performance questions are 30

3.5 Communicating Synchronous Finite State Machines

always decidable since each state can, in theory, be examined in nite time. In practice, however, this may be prohibitively expensive, and thus formal veri cation techniques based on interacting FSMs require various forms of (non-trivial and non-automatable) abstraction in order to be kept manageable 38, 45]. A traditional FSM consists of: a set of input symbols (the Cartesian product of the sets of values of the input signals), a set of output symbols (the Cartesian product of the sets of values of the output signals), a nite set of states with a distinguished initial state, an output function mapping input symbols and states to output symbols, and a next-state function mapping input symbols and states to (next) states. The input to such a machine is a sequence of input symbols, and the output is a sequence of output symbols. The model is synchronous (i.e., all signals have the same tags), and hence input and output symbols are well de ned (they correspond to the set of events with a given tag). It is also semantically identical to that of previous section. However, there are enough syntactic di erences to warrant a separate treatment (see 28, 11] for a discussion of possible mappings between the two). Traditional FSMs are good for modeling sequential behavior, but are problematic for modeling system with concurrency or large memories, because of the state explosion problem. Every global state of a concurrent system must be represented individually, even when interleaving of independent actions may give rise to an exponential number of states. Similarly, a memory has as many states as the number of values that can be stored at each location raised to the power of the number of locations. The number of states alone is not always a good indication of complexity, but it often has a strong correlation. Harel advocated the use of three major mechanisms that reduce the size (and hence the visual complexity) of nite automata for modeling practical systems 31]. The rst one is hierarchy, in which a state can represent an enclosed state machine. That is, being in a particular state a has the interpretation that the state machine enclosed by a is active. Equivalently, being in state a means that the machine is in one of the states enclosed by a. Under the latter interpretation, the states of a are called \or states." Or states can exponentially reduce the complexity (the number of states) required to represent a system. They compactly describe the notion of preemption (a high-priority event suspending or \killing" a lower priority task), that is fundamental in embedded control applications. 31

The second mechanism is concurrency. Two or more state machines are viewed as being simultaneously active. Since the system is in one state of each parallel state machine simultaneously, these are sometimes called \and states." They also provide a potential exponential reduction in the size of the system representation. The third mechanism is non-determinism. While often non-determinism is simply the result of an imprecise (maybe erroneous) speci cation, it can be an extremely powerful mechanism to reduce the complexity of a system model by abstraction. This abstraction can either be due to the fact that the exact functionality must still be de ned, or that it is irrelevant to the properties currently considered of interest. E.g., during veri cation of a given system component, other components can be modeled as non-deterministic entities to compactly constrain the overall behavior. A system component can also be described non-deterministically to permit some optimization during the implementation phase. Non-determinism can also provide an exponential reduction in complexity. Note that non-determinism can be divided into and-non-determinism and or-non-determinism. In the rst, the resolution of the non-determinism executes all possibilities, while in the second, resolution chooses just one. Andnon-determinism is equivalent to hierarchy. These three mechanisms have been shown in 26] to cooperate synergistically and orthogonally, to provide a potential triple exponential reduction in the size of the representation with respect to a single, at deterministic FSM8 . Harel's Statecharts model uses a synchronous concurrency model (also called synchronous composition). The set of tags is a totally ordered countable set that denotes a global \clock" for the system. The events on signals are either produced by state transitions or inputs. Events at a tick of the clock can trigger state transitions in other parallel state machines at the same clock. Unfortunately, Harel left open some questions about the semantics of causality loops and chains of instantaneous (same tick) events, triggering a urry of activity in the community that has resulted in at least twenty variants of Statecharts 65]. A model that is closely related to FSMs is Finite Automata. FAs emphasize the acceptance or rejection of a sequence of inputs rather than the sequence of output symbols produced in response to a sequence of input symbols. Most notions, such as composition and so on, can be naturally extended from one model to the other. FAs without accepting conditions are also called Labeled Transition Systems in the literature.

8 The exact claim in 26] was that and-non-determinism (in which all non-deterministic choices must be successful), rather than hierarchical states, was the third source of exponential reduction together with \or" type non-determinism and concurrency. Hierarchical states, on the other hand, were shown in that paper to be able to simulate \and" non-determinism with only a polynomial increase in size.

32

Synchronous FSMs, as described above, have a clear and deterministic composition mechanism that makes them relatively easy to understand, synthesize and verify. Of course, there is also a signi cant drawback: deciding when composition is well-de ned (loosely speaking, there are no combinational loops) has a high computational complexity. Moreover, for many applications, the tight coordination implied by the synchronous model is inappropriate. In particular, it is very di cult to keep a tight synchronization between heterogeneous components of an embedded system, since the pace of a synchronous system is dictated by its slowest component. In response to this, a number of more loosely coupled asynchronous FSM models have evolved, including CSP 33], CCS 46], behavioral FSMs 61], SDL process networks 61], and codesign FSMs 21]. In this section we focus on process algebraic models that constitute the semantical foundation of the Occam and Lotos 64] languages9: Communicating Sequential Processes 33] and the related Calculus of Communicating Systems 46]. In the following we discuss only the control aspect of CSP and CCS, and ignore the fact that their processes can also manipulate data via assignments, tests and so on. We also do not consider recursion, that can be de ned in the process algebra but has limited interest (except for tail recursion, that de nes looping) in the context of embedded systems. The behavior of each process is modeled by a Labeled Transition System (only nite LTSs are of interest in embedded system design, for obvious reasons). Arcs in the transition system are labeled with signal names, and the state transition activity imposes a total order on the signals of each process. Communication is based on rendezvous. That is, two LTSs may share a signal, thus imposing that all the events of that signal must occur in both processes (\at the same time", if we interpret tags as time). Finally, process algebrae generally imply a completely interleaved view of concurrent actions, meaning that no two events may have the same tag . Concurrent (i.e., independent) events occur in all possible interleaving in the LTS. No two events may have have the same tag, and hence process algebrae are an inherently asynchronous model. Note that a single LTS is an interleaved asynchronous model, while multiple LTSs communicating via rendezvous (and, equivalently, Petri nets in which at most one token can reside in each place in each reachable marking) are a partially ordered asynchronous model. As mentioned above, the rich theory of regions 48, 23] can be used to freely move between the two classes of models. The result of process composition using this communication mechanism is another LTS, thus resulting in a hierarchical compositional model10. Composi9 Ada also uses rendezvous, although the implementation is stylistically quite di erent, using remote procedure calls rather than more elementary communication primitives. 10 Compositionality means that two or more communicating processes can be viewed as a

3.6 Process algebrae

33

r I a err (a) τ II 1

r 2 a err 3 (b)

r 1,I a τ err 3,I (c) 1,II r 2,II 1,I

r 2,II a err 3,I (d)

Figure 6: Example of Labeled Transition Systems and rendezvous communication tionality is very important for proving properties of the system in a hierarchical fashion. This property is also true of communicating synchronous Finite State Machines, but not of data ow networks (i.e., a data ow network is di erent from an actor). Let us consider a simple case of an interface with error detection. The LTSs specifying the protocols followed by the two partners are shown in Figure 6.(ab). 1. The sender has two states, rst sending a request on signal R, then waiting for either an acknowledgment of correct reception on signal A, or an error indication on signal E . 2. The receiver has a similar behavior, but in case of error, it requires one internal action (labeled ) to resynchronize, and hence it has a third state. The composed LTS using the rendezvous mechanism is shown in Figure 6.(c). Note how the state space of the composition is the product of the two state spaces, and the two LTSs synchronize on common edge labels. For the sake of comparison, Figure 6.(d) shows the synchronous composition of the same two LTSs. Note how in case of error, the receiver waits for one clock tick, and hence becomes de-synchronized with the transmitter, thus leading to a deadlock11. Rendezvous-based models of computation are sometimes called synchronous in the literature. However, by the de nition we have given, they are not synsingle process, that can in turn be used as a unit and composed with others. 11 Of course, the fact that synchronous composition deadlocks while asynchronous composition does not is just a coincidence. It is easy to construct an example where the converse can happen.

34

chronous. Events are partially ordered, not totally ordered, with rendezvous points imposing the partial ordering constraints. SDL 61] is a language for speci cation, simulation and design of telecommunication protocols. Its underlying semantical model12 is based on a process network. Each process is an FSM, and communication is via one unbounded FIFO queue per process. If we ignore the ability of a process to manipulate its input queue, the MOC is roughly equivalent to DE, with the restriction that the FSM can only read one event at a time. SDL networks have a basic implementability problem, since both the size of the queues and the topology of the network can change at run time. (Processes can be created on the y, and signals can be routed dynamically based on process identi ers.) Hence they either require a software implementation based on a Real-Time Operating System with dynamic memory allocation and task instantiation, or require the designer to pre-size queues and pre-instantiate all processes.

3.7 SDL process networks

3.8 Timed Automata

Synchronous and asynchronous Finite State Machines cannot reason easily about time, since in the best case (the synchronous one) time must be represented by counting clock ticks. This may cause a state explosion, and has been proven to be an inadequate abstraction of reality unless special care is taken 19]. For this reason, Alur and Dill 2] have proposed explicitly introducing time as a continuous quantity in the Timed Automata MOC. A Timed Automaton (TA) is a special case of hybrid systems, which are described in the next section. It is su ciently restricted, so that most properties become decidable. A TA is a Finite Automaton (FA) plus a set of clocks . The state of the TA is the state of the FA together with a real valuation of the clocks. A transition of the TA is labeled with a symbol (from the FA alphabet) and a Boolean formula over atomic propositions comparing clocks with integers. The transition can also reset some clocks to zero. While the state space of a TA is clearly in nite, a key result by Alur and Dill shows that it admits a nite state representation, by means of a partition into equivalence classes. Basically, 2] showed that the exact value of a clock does not matter after it grows beyond the largest constant with which it can be compared in any transition label. This imposes an equivalence relation on those portions of the state space that grow towards in nity. Moreover, since comparisons involve only integers, one can also partition the remaining part of

12 As usual, we focus on the control and communication aspects over the data computations, which are commonly speci ed with an imperative host language, in addition to a more formal and less practical treatment based on Abstract Data Types.

35

the space into a nite set of equivalence classes (called regions), that admit a normal form representation (computed via an all-pair shortest path algorithm). This is a very signi cant contribution, however it has shown only limited practical applicability so far because the state explosion problem is even more severe than in the communicating FSM case. Good generally applicable abstraction techniques are only beginning to be developed for TAs. A hybrid system is a Finite Automaton in which each state is associated with a set of di erential equations, and transitions occur when inequalities over the continuous variables of the di erential equations are satis ed. Hybrid systems are a powerful mechanism for modeling non-linear dynamic systems, and thus are becoming an essential tool in control theory. However, they are clearly Turing-equivalent, and hence too powerful, in almost all of their incarnations, with the notable exceptions of Timed Automata described above. It is likely that they will play an ever increasing role in embedded system design due to the growing need to raise the level of abstraction, but it is di cult to give them a complete and fair treatment in this brief overview, and we refer the interested reader to 32]. In the TSM, there are two possible views of hybrid systems (and hence of TAs). 1. A hybrid system (FA plus di erential equations) can be modeled as a single TSM process. This provides an easy mechanism for composing hybrid systems. Signal tags in this case are order-isomorphic with the real numbers, but tags in which a transition of the automaton can occur can be only discrete. 2. A hybrid system can be modeled as a set of TSM processes. In this case we have two components for each hybrid system: one process, whose signal tags can only be discrete, represents the automaton, and multiplexes the hybrid system outputs between a set of processes, each behaving as a set of di erential equations.

3.9 Hybrid systems

4 Codesign Finite State Machines

Codesign Finite State Machines (CFSMs) are the underlying MOC of the POembedded system design environment 21, 6]. We describe them at length because, as we will argue later, they combine interesting aspects from several other MOCs, while preserving both formality and e ciency in implementation. As we pointed out above, one of the most important properties of an MOC is synchronicity or asynchronicity. We wish to summarize our views on this topic to motivate the introduction of yet another MOC.

LIS

36

Synchrony and asynchrony represent two fundamentally di erent views of time. That is, synchrony uses the notions of zero and in nite time, while asynchrony uses non-zero nite (and typically bounded) time. Both synchrony and asynchrony have appeared a number of times in our previous descriptions of various models of computation. In this section, we summarize our previous presentations of synchrony and asynchrony, and consider the di erences in the behaviors produced under each model. As usual, we consider a system of processes interacting through events.

4.1 Synchrony and asynchrony

reads inputs, computes, and produces outputs simultaneously. That is, all the synchronous events (both inputs and outputs) happen simultaneously, implying zero-delay calculations. In between clock ticks, an \in nite" time passes. Of course, no calculations happen in zero time in practice, nor does one wait an in nite amount of time between ticks (it is normally nite but unspeci ed). In practice, the computation times are much smaller than the clock rate, and thus can be considered to be zero with respect to the reaction times of the environment. The very desirable feature of designs implemented as synchronous systems with no cyclic dependencies among values of events with the same tags, is that the behavior of the implementation is not dependent upon the timing of the signals, thus simplifying tremendously the veri cation task.

4.1.1 Synchrony Basic operation: At each clock tick (i.e., tag of its signals), each module

Triggering and Ordering: All modules are triggered to compute at every clock tick. At a tick, there is no ordering of reading of inputs, computations, or writing of outputs. However, an ordering can be imposed in addition with the concept of delta steps (delays). A delta step (delay), as previously mentioned, is the (zero) time that passes between events at the same clock tick and which serves simply to order the events.

inputs. A well-designed synchronous system will have a unique solution (assignment to all signals) at each clock tick, though the corresponding models of computation, as well as many synchronous languages or speci cation methods, allow the designer to specify systems that do not have this property (see, e.g, 65]). We recall that the presence of cyclic dependencies among values of events with the same tag are responsible for this di culty. It is the domain of the language and its semantical interpretation to verify whether a unique solution exists. Synchronous systems that have a unique solution, have a \single" nite state machine equivalent even though they consist of several interconnected components, and thus can be analyzed and veri ed with e cient techniques. 37

System solution: The system solution is the output reaction to a set of

Implementation cost: Adherence to the synchronous assumption, that is, a process computes in negligible time compared to its environment, is a property that must be veri ed or enforced on the nal design, and which may be expensive to implement. The assumption is checked on the nal implementation. For hardware, one must ensure that the clock period is higher than the maximum possible computation time for a synchronous block; this may imply a clock rate that is much slower than might otherwise be achieved. For software, one must ensure that an invoked process is allowed to complete before another process or the operating system changes its inputs. 4.1.2 Asynchrony Basic operation: Asynchronous events always have a non-zero amount of

time between them: it is impossible to specify that two events happen simultaneously in a truly asynchronous system (as in real life : : : ). An individual process can run whenever it has a change on its inputs, and it may take an arbitrary time (that is typically bounded) to complete its computation.

triggered to run) when it has inputs that have changed. However, among the triggered modules, there is no a priori ordering of processes. One may later be imposed by a scheduling algorithm, but this is part of the implementation choice.

Triggering and Ordering: A module is only triggered to run (and always

System solution: There is strong dependency of the solution from input signals and their timing. Thus, asynchronous systems are much more di cult to analyze. In addition, in a practical implementation or a model thereof, some events may appear to happen simultaneously. In practice it may be di cult and expensive to maintain the total ordering. If the actual order of these seemingly simultaneous events is not preserved, any order may be used possibly resulting in multiple behaviors. This is no longer an asynchronous model but a discreteevent model that has no guarantee of uniqueness of the solution because of the possible cyclic dependency of values of events with the same tags. It is this practical aspect that has misled many when assessing the properties of asynchronous systems, loading asynchronous systems with problems that are typical of discrete-event systems. Implementation cost: Asynchronous implementations are usually chosen

when the cost, particularly in terms of computation time, is too high for a synchronous solution. The exibility provided by an asynchronous implementation implies that di erent parts of the same system (or the same system under di erent inputs) can operate at quite di erent rates, only communicating at particular check-points in the computation. For system design, it is usually 38

imperative to have an asynchronous model at the highest level of communication. On the other hand, analysis of the behavior of designs implemented as asynchronous systems has to take into consideration the timing of the signals and, hence, is much more complex than the analysis of synchronous systems. This is the reason why much research on asynchronous system has been devoted to implementations that are more or less insensitive to \internal" delays, thus retaining the most desirable property of synchronous systems without paying the full penalty implied in a synchronous implementation. An ideal MOC for system design should combine the advantages of veri ability in synchrony and exibility in asynchrony in a globally asynchronous, locally synchronous (GALS) model. It is important to be explicit about where the boundary is between synchrony and asynchrony, because the behavior of the two, clearly, is very di erent. The di erences can be illustrated simply in terms of event bu ering and timing of event reading/writing. In an asynchronous implementation, there is typically a need for an explicit bu ering mechanism for the events, since it is not known when a module will run and hence read its inputs, and since di erent modules will run at di erent times and use the same input at di erent times. For synchrony, inputs are all read once at the beginning of a computation, so one global copy of each event value su ces and this one copy is cleared at the end of each tick. Thus, a synchronous communication transmits all events simultaneously and in zero time with no bu ering; every module is guaranteed to see the same set of events at each clock tick. An asynchronous communication transmits events when ready and through bu ers; each module sees its own stream of inputs which depends on the global scheduling. Many programs will behave the same for an asynchronous or synchronous implementation, and such systems are typically more tolerant to implementation uctuations. One can program in a style that is more robust with respect to these di erences, by, for example Never assuming or waiting for simultaneous (synchronous) events. Since simultaneity is nearly impossible to guarantee, it is more robust to wait for the occurrence of two events rather than the simultaneous occurrence of them. Never programming with a global timing, e.g. a global clock tick, in mind. Synchronous languages have mechanisms for allowing a clock tick to pass, and thus for counting clock ticks and waiting for a certain amount of this arti cial time to pass. Asynchronous systems of course do not have such a speci c notion of time, so the same style of programming with an asynchronous model interpretation (in which a clock tick usually forces an ordering rather than referring to a time) will produce di erent behavior. 39

4.1.3 Combining Synchrony and Asynchrony

One may certainly use these programming techniques within a synchronous portion of the design. At the system level, however, a time-intolerant style of programming and thinking about the behavior of a design should be employed. Our CFSM model re ects these views and was strongly motivated by the need of combining synchronous and asynchronous behavior where it made most sense. Each CFSM is an extended FSM, where the extensions add support for data handling and asynchronous communication. In particular, a CFSM has a nite state machine part that contains a set of inputs, outputs, and states, a transition relation, and an output relation. a data computation part in the form of references in the transition relation to external, instantaneous (combinational) functions. a locally synchronous behavior: each CFSM executes a transition by producing a single output reaction based on a single, snap-shot input assignment in zero time. This is synchronous from its own perspective. a globally asynchronous behavior: each CFSM reads inputs, executes a transition, and produces outputs in an unbounded but nite amount of time as seen by the rest of the system. This is asynchronous interaction from the system perspective. This semantics, along with a scheduling mechanism to coordinate the CFSMs, provides a GALS communication model: Globally (at the system level) Asynchronous and Locally (at the CFSM level) Synchronous.

4.2 CFSM Overview

4.3.1 Signals

CFSM

4.3 Communication Primitives

s, as TSM processes, communicate through signals, which carry information in the form of events. They may function as inputs, outputs, or state signals. A signal is communicated between two CFSMs via a connection (singleinput, single-output communication process) that has an associated input bu er (or 1-place bu er), which contains one memory element for the event (event bu er) and one for the data (data bu er). The event is emitted (produced) by a sender CFSM setting the event bu er to 1. It may be detected and consumed by a receiver CFSM. It is detected by reading the event bu er; it is consumed by setting the bu er to 0. 40

A signal is therefore present if it has been emitted and not yet consumed. In the tagged signal model, this means that the input signal of the connection has had an event with a tag larger than the largest tag of the output signal. The data may be written by a sender and read by a receiver. Reading and writing is done on the data bu er on the connection between the sender and the receiver. A control signal carries only event information, i.e., it may only be emitted and detected/consumed and its value is irrelevant. A data signal carries only data information, i.e., it may only be read and written. An input signal can only be detected/consumed and read (depending on its status as a signal, control signal, or data signal). A (possibly incomplete) set of values for the input signals of a CFSM is termed input assignment, a set of input values read by a CFSM at a particular time is termed captured input assignment, and an input assignment with at least one present event is termed input stimulus. An output signal can only be emitted and written. A (possibly incomplete) set of values for the outputs of a CFSM is termed output reaction. A state signal is an internal input/output data signal; it may be written and subsequently read by its CFSM. A state is a set of values for the state signals. A set of states may be given by a subset of state values. States are implicitly represented by the state signals and hence may be encoded or symbolic. The state signals could be considered part of the input and output signal sets, and it is only for the exposition that they are separated: discussion of scheduling and runnable CFSMs is facilitated by identifying an input assignment that triggers the CFSM separately from its state. Where the type is unimportant, we may refer to any of the basic signal types (signal, control signal, data signal, input signal, output signal, state signal) simply as signal. As will be seen in the behavior sections to follow, CFSMs initiate communication through events. The input events of a CFSM determine when it may react. That is, the model forbids a CFSM to react unless it has at least one input event present (except for the initial reaction, described in the functional behavior section). Without this restriction, a global clock would be required to execute the CFSMs at regular intervals, and this clock would in fact be a triggering input for all CFSMs. This would clearly imply a more costly implementation. A CFSM can trigger itself by emitting an output and detecting that same signal in the next execution. A CFSM with at least one present input event is termed runnable. For CFSM A to send a signal S to B, A writes the data of S and then emits its event. This order ensures that the data is ready when the event is communicated. B is scheduled, sees the event (which is its stimulus), reads the corresponding data, and reacts. Pure data signals will only be read and written by a CFSM that has already been triggered by the presence of another input event. 41

A net is a set of connections on the same output signal, i.e., it is associated with a single sender and at least one receiver (in the TSM, it is a set of connection processes with the same input). There is an input bu er (TSM connection) for each receiver on a net, hence the communication mechanism is multi-cast: a sender communicates a signal to several receivers with a single emission, and each receiver has a private copy of the communicated signal. Each CFSM can thus independently detect/consume and read its inputs. A network is a set of CFSMs and nets. The behavior of the network (and even of a single CFSM) depends on both the individual behavior, and that of the global system. In the mathematical model, the system is composed of CFSMs and a scheduling mechanism coordinating them. It can be implemented as: a set of CFSMs in software (e.g., C), a compiler, an operating system, and a microprocessor (the software domain), a set of CFSMs in hardware (e.g., gates mapped to an FPGA), a hardware initialization scheme, and a clocking scheme (the hardware domain), and the interface between them (e.g., a polling or interrupt scheme to pass events from hardware CFSMs to software ones via the RTOS, a memorymapped scheme to pass events from software to hardware). Thus the scheduling mechanism in the model may take several forms in the implementation: a simple RTOS scheduler for software on a single processor and concurrent execution for hardware, or a set of RTOSs on a heterogeneous multi-processor for software and a set of scheduling FSMs for hardware. The CFSM model does not require any coordination between these schedulers in order to guarantee correct behavior, apart from an implementation of the event delivery mechanism (the interface). Explicit or implicit coordination is required only in order to satisfy timing constraints, which in turn may guarantee an ordering of events and/or a particular functional behavior. In the CFSM model, a global \scheduler" controls the interaction of the CFSMs, and invokes each appropriately during execution of the design. The system output will depend on the functional and timing behavior of the individual CFSMs, and the functional and timing behavior of their ensemble. The scheduler operates by continually deciding which CFSMs can be run, and calling them to be executed. Each CFSM is either idle (waiting for input events or waiting to be run by the scheduler), or executing (generating a single reaction). During an execution, a CFSM reads its inputs, performs a computation, and possibly changes state and writes its outputs. 42

4.3.2

CFSM

Networks

4.4 Timing Behavior

The mathematical model places few restrictions on the timing of an execution. Each CFSM execution can be associated with a single transition point, ti, in time. The model dictates that it is at this point that the CFSM begins reacting: reading inputs, computing, changing state, and writing outputs. Since the reaction time is unbounded, one cannot say exactly at which time a particular input (event or data) is read, at which time that input had previously been written, or at which time a particular output is written. There are, however, some restrictions. For each execution, each input signal is read at most once, each input event is cleared at every execution, and there is a partial order on the reading and writing of signals. Since the data value of a signal (with an event and data part) only has meaning when that signal is present, the model dictates that the event is read before the data. Similarly for the outputs, the data is written before the event, so that it is valid at the time the event is emitted. This means that for transition point ti , an input may be read at any time between ti and ti+1 (but not later, because that would correspond to transition point ti+1 ), the event that is read may have occurred at any time between ti?1 and ti+1 , the data that is read may have been written at any time between t0 and ti+1 , and the outputs are written at some time between ti and ti+1 . After reading an input, its value may be changed by the sender before ti+1 , but the receiver reacts to the captured input and the new value is not read until the next reaction. This exibility in timing can have non-intuitive behavior. Example: Event/data separation. Suppose a sender S writes data value v1 at t1 for signal X and emits it at t2 . Let this emission be e1, so the pair is (e1, v1). This is illustrated in Figure 7. A receiver R at ti sometime later reads the event, but takes longer than expected to read its corresponding value v1. S communicates X again: value v2 then emission of X, for pair (e2, v2). R now reads the value for X, and reads v2. The captured input for the R is thus the pair (e1, v2), which was not the pair intended by S. Furthermore, data v1 has been lost forever, even though it was sent with an event to signal its presence. Problems such as this can easily be resolved by requiring the appropriate level of atomicity in the model and in the implementation, i.e., by restricting some parts of the communication to take place simultaneously and instantaneously { as a single entity. In the CFSM model, the requirement is simply that the input events are read atomically. At each ti, a CFSM reads its input events without interruption, and without those events being overwritten by the sending CFSMs. This is easily implemented in software by reading a bit-vector of input events in one instruction, and in hardware by clocking all 43

X:=v1 Sender S

emit (X) (e1:=1)

X:=v2

emit (X) (e2:=1) time

Receiver R

Read events

Read value X

ti?1

t1

t2

ti

t3

t4

ti+1

Receiver R captures (e1, v2)

Figure 7: Event/data separation. s together, with a separate read phase and a compute/write phase (per clock cycle). Atomicity of input event reading implies an implementation that retains much of the exibility required for e ciency, while mitigating the worst of the synchronicity problems. It should be clear that allowing input event and data reading and output event and data writing to happen at completely arbitrary times leads to behavior with very di cult to predict and prescribe results. Given the event-based communication of CFSMs, atomicity of input event reading is a natural means of ensuring some predictability: a receiver CFSM is guaranteed to see a snapshot of input events that are simultaneously present at some point in real time. Additional constraints, if necessary, can be imposed to ensure that the values subsequently read are meaningful. These constraints will vary considerably depending on the implementation chosen and the design constraints. Example: no atomicity of data reading. Consider a sender S and two receivers R1 and R2, as illustrated in gure 8. S is sending the value of signals X and Y. Both X and Y are currently 4 and are changing to 5. R1 reads X at t1 and Y at t6 . R2 reads X at t3 and Y at t4. S changes X at t2 and Y at t5. R1 therefore captures X = 4 and Y = 5 while R2 captures X = 5 and Y = 4. Not only do they capture di erent input assignments, but R1 captures a set of values that never occurs simultaneously. Note that if X and Y were (control/data) signals, they would be sent with events as well, and they would both be changed to 5 before the events were emitted and hence before the receivers can read them. However, those new values can be overwritten by the sender if the receiver doesn't read fast enough, leading to the separation of the event/data pair as illustrated in the previous example. Example: atomicity of event reading. Now consider a system with a

CFSM

44

X:=4 Y:=4 Sender S X:=5 Y:=5

Receiver R1

Read X

Read Y time Read X Read Y

Receiver R2

t1

t2

t3

t4

t5

t6

Receiver R1: X = 4, Y = 5 Receiver R2: X = 5, Y = 4

Figure 8: No atomicity of data value reading. sender S and three receivers R1, R2, and R3 as illustrated in Figure 9. S emits control signals X and Y at times t2 and t4 . R1 reads at t1 and sees both absent. R2 reads at t3 and sees only X present. R3 reads at t5 and sees both X and Y present. Though each has a di erent captured input assignment, each sees an input assignment that occurs at some point in real time. The functional behavior of a CFSM at each execution is determined by the speci ed transition relation (TR). This relation is a set of tuples (input set, previous state, output set, next state) where input set is an input assignment, previous state is a state, output set is an output assignment, and next state the next state. Each tuple of the TR represents a speci ed transition of the machine, and the set of tuples is the speci ed behavior of the machine. A transition in which the input set includes an input stimulus is termed a valid transition. At each execution, a CFSM 1. Reads an input assignment. 2. Looks for a transition Transition = ( input set, previous state, output set next state) such that the read input includes input set and the present state of the CFSM matches present state (hence the absence of an input from input set means \don't care about that input"). 45

4.5 Functional Behavior

emit (X) Sender S

emit (Y)

Read X Receiver R1 time

Receiver R2

Read X

Receiver R3

Read X

t1

t2

t3

t4

t5

Receiver R1 : X, Y absent Receiver R2 : X present, Y absent Receiver R3 : X, Y present

Figure 9: Atomicity of event reading. 3. If Transition is found, it is executed by (a) consuming the inputs (setting input event bu ers to 0) (b) making the state transition to next state (c) writing the new output events in output set (hence absence from output set means \don't emit/don't modify"). 4. If Transition is not found, the CFSM consumes no inputs, makes no state change, and writes no outputs. The last case, in which no matching transition is found, is known as the empty execution. If this can happen for some input stimulus, the transition relation is incomplete; otherwise it is complete. For software, this is precisely the same behavior that would be produced if this CFSM had not been scheduled by the RTOS. If several transitions match, the CFSM is non-deterministic and the execution can perform any of the matching transitions. For the implementation, all CFSMs must be deterministic in order to simulate and synthesize the behavior. Non-determinism can be used at the initial stages of a design in order to model partial speci cations, later to be re ned into deterministic CFSMs. A trivial transition is one in which no output events are emitted, no output values are changed, and no state change is e ected, but inputs are consumed. It 46

e ectively discards the current input assignment and waits for a new one. Trivial transitions are speci ed in the TR like any other transition, with output set empty (or leaving state variables unchanged). Each state variable may have a designated set of reset values (or initial values) that are speci ed with the transition relation. A set of reset values, one for each state variable, is a reset state (or initial state). If there are several reset values for a state variable, there are several reset states. This represents a non-deterministic starting condition which must be resolved before synthesis or simulation can be performed. The initial transition(s) is a special transition(s) where the present state part is equal to the reset state. Moreover, this transition is allowed to not have any input events present in the input assignment. (Recall that for all other transitions, at least one input event is required to trigger the CFSM.) The initial transition(s) may be speci ed for a CFSM, but is not required. There may be several possible initial transitions depending on the initial state(s) and the values of the corresponding input assignments. If there is non-determinism, again, it must be resolved before synthesis and simulation. We can now classify CFSMs along the same lines that were used for the other formal models. CFSMs are an asynchronous Extended FSM model, that is di erent from CSP and CCS because communication is not via rendezvous but via bounded (1-deep) non-blocking bu ers, and di erent from SDL since queues are bounded and the process network topology is xed. Moreover, each CFSM can be modeled with an LTS in which each edge label can involve presence and absence tests of several signals, while in CSP, CCS, and SDL each label consists of a single symbol. Signals are distinguished among inputs and outputs. Transitions, unlike data ow networks, can be conditioned to the absence of an event over a signal. Hence CFSMs are not continuous in Kahn's sense 35] (arrival of two events in di erent orders may change the behavior). The semantics of a CFSM network is de ned based on a global explicit notion of time (imposing a total ordering of events). Thus CFSMs can be formally considered as synchronous with relaxed timing . I.e., while a global consistent state of the signals is required in order to perform a transition, no relationship is required between the tags of input events involved in a given transition, nor between those of its output events. There is only a partial order relationship between input and output events of the transition (inputs must have tags smaller than outputs). Finite bu ering without blocking write implies, as mentioned above, that events can be overwritten , if the sending end is faster than the receiving end. 47

4.6 CFSMs and process networks

A

i

i1 err i2 o

C

B

Figure 10: Example of CFSM network. This sort of \deadline violation" in the CFSM context may or may not be a problem, depending both on the application and on the type of event. The designer can make sure that \critical" events are never lost: either by providing an explicit handshaking mechanism, built by means of pairs or sets of signals, between the CFSMs, or by using scheduling techniques that ensure that no such loss can ever occur 5]. In this sections we provide a few examples of what it means to specify a \behavior" with a relaxed timing model such as CFSMs, and we discuss the notion of behavior equivalence classes . Consider a simple case of three CFSMs as shown in gure 10. These CFSMs specify an \almost data ow" behavior. CFSMs A and B take the same input stream i, and perform two di erent kinds of (unimportant) processing on it, by producing an output event for every input event. CFSM C takes an event from each input i1 and i2 and produces either an event on its o output if there is no error (e.g., its inputs are within a speci ed range) or an event on its err output, if some problem in the input stream or the CFSM state is detected. The err signal causes A and B to perform some recovery action (e.g., realign their state variables). The intuitive behavior speci ed by these three CFSMs is, in the designer's eyes, the same regardless of the scheduling in time of CFSM transitions, as long as: 48

4.7 Examples of CFSM behaviors

1. no events are lost (they are all \critical" in this case), and 2. (possibly) some latency constraint is satis ed (e.g., o may be needed earlier than the next external input arrival). This means that, given the choice of possible timed executions of these CFSMs, they are partitioned into a set of equivalence classes . Let us consider, for the sake of simplicity, only reaction times and no other scheduling constraints (each CFSM is allocated its own hardware resource or processor). Let us assume that every CFSM has the same reaction time of nr time units if there are no errors, and of 2nr in case of errors (C when it emits err , A and B when they detect err ). Let us also assume that input events arrive at a regular rate of ni time units, and that there are only \no-missed-event" constraints and no other no latency constraints. In this case, we can consider the following equivalence classes with respect to the above mentioned intuitive behavior: 1. Zero-delay executions: nr = 0. These are logically inconsistent (noncausal in the Esterel terminology 11]), since if C detects an error, A and B should instantaneously react and produce di erent outputs (conceivably without the error conditions). This is clearly absurd. 2. Executions in which the execution delay of A, B and C is larger than the inter-arrival time of inputs: ni < nr . These executions clearly do not satisfy the intuitive behavior requirements listed above. 3. Executions in which the delay of A, B and C is smaller than the interarrival times of external inputs, but larger than half of that: ni =2 < nr ni . These executions handle correctly the normal ow of data, but \miss" an input in the case of error. This happens because the execution time of C is too long and causes it to miss the rst input events after the error. Also A and B are too slow and miss an input event when recovering. 4. Executions in which the delay of A, B and C is smaller than half of the inter-arrival times of external inputs: nr ni =2. In this case, no event is lost. If errors are infrequent enough, the designer may want to consider the two last classes to be equally good, and accept the cheapest one. On the other hand, if errors are frequent, or if every single input really matters (there is zero redundancy in the input stream), only the last, conceivably most expensive equivalence class is acceptable. Note how the notion of equivalence classes can be applied to analyze also executions in which a scheduler coordinates CFSMs by enforcing mutual exclusion constraints (this is an appropriate model, e.g., for a single-processor implementation). In that case event loss can occur due to 49

Timing constraints. E.g., ni < 3nr would imply that an execution falls into the second class above, and misses deadlines due to excessive processor occupation. \Incorrect" scheduling. E.g. if 3nr ni < 4nr and the scheduler activates the CFSMs in the xed order ACBC. In that case, the rst activation of C is valid, since it has an input stimulus, but redundant, since it always results in an empty execution (assumed with the same nr delay), and causes the system to miss a deadline even during \normal operation". On the other hand, another valid schedule ABC will not miss deadlines in normal operation. Obviously this de nition of \equivalence classes" between behaviors of a CFSM network is very application-dependent, and as such di cult to formalize. Here we can suggest only a few criteria that could be used for this purpose, such as: 1. Equality between streams of values produced at some output by two di erent timed behaviors of the same network, given the same stream of values on the inputs (\data ow" equivalence). 2. Compatibility with a given partial ordering between events (\Petri net" equivalence). 3. No missed critical events, possibly quali ed as, e.g., \no missed events except for the rst n events after abnormal event x" (\quasi-data ow" equivalence). 4. Equality of input-output sequences, possibly modulo reordering of \concurrent events", with respect to a completely deterministic reference speci cation (\golden model" equivalence). 5. Equality of input-output sequences modulo ltering by some testbed entities that model the external, physical system constraints (\ ltered" equivalence). While we are still far from a formalization of these criteria, we believe that the richness of the CFSM model stems, among other factors, from the ability to exploit all these sorts of equivalence while, for example, data ow networks can exploit only one (data ow equivalence). We will further elaborate on this in the next section.

5 Conclusions

The relative advantages and disadvantages of the various MOCs have been described in the previous sections. We are still far from having a single agreedupon standard MOC that is suitable for all types of embedded system designs. 50

Some authors ( 17]) advocate heterogeneity at the MOC level as an essential requirement of embedded systems. However, based on the discussions above, we can identify a new MOC that is expressive enough to capture most practical embedded systems, and formal to permit e cient veri cation and synthesis of some special cases. This model is that of CFSMs with initially unbounded FIFO bu ers . Bounds on bu ers (essential for implementability) are imposed by re nement , exactly as timing information is re ned in the original CFSM model. The motivations for this proposal are as follows. 1. Local synchrony: concentrating the control inside relatively large atomic synchronous entities helps the designer to better understand the overall coordination. Models such as Colored Petri Nets, which view control at a ner level of granularity, are di cult to use for large realistic designs. 2. Global asynchrony: breaking synchronicity helps resolving composition problems and mapping to a heterogeneous architecture. Synchronous models, as argued previously, cannot handle multi-rate e ciently, especially when the rates of di erent signals are totally uncorrelated. This is essentially due to the need to always consider all signals, including those that are not present in most clock cycles. 3. Unbounded bu ers: leaving bu ers unbounded at the outset o ers opportunities to perform static and quasi-static scheduling whenever possible ( 40, 18, 47]). As a result, some bu ers become statically sized, as determined by the static schedule of reader and writer processes. The designer sizes the remaining bu ers to ensure implementability, and to use simulation, formal veri cation or Real-time scheduling 4, 5] to validate the design in presence of nite FIFO bu ers. The use of lossy bu ers (i.e., with non-blocking write) is somewhat arguable, because there are cases in which loss is essential (in general, any time there are tight timing constraints, that make some of the signals irrelevant under some conditions, e.g., an emergency), and there are cases in which loss is problematic (in general, any time one would like to model data ow computations or any other blocking communication mechanism, such as, e.g., Remote Procedure Call). Our choice is to keep bu ers lossy in the formal model , and give the designer tools to verify a priori if loss can occur , as well as to enforce no loss for some bu ers in the implementation . In general this can be enforced at an acceptable cost only under some speci c conditions, e.g., that the lossless bu er must be local to a processor, or that the communicating processes must be statically scheduled with respect to each other. 51

The resulting model combines interesting properties of the main MOCs seen above, while still keeping a strong link to veri ability and implementability. In particular: At the initial, untimed level it describes a partial ordering between signal tags, and hence captures a whole class of possible implementations on a variety of architectural options. These options include software on multiple processors, pipelined hardware, and so on. It keeps computation (in the FSM), communication (in the bu ers) and timing (in the architectural mapping) as separate as possible. After architectural mapping, it becomes essentially a Discrete Event model, and thus lends itself to performance and power consumption analysis, in order to evaluate architectural trade-o s. A subset, such that CFSMs have a deterministic behavior (i.e., behave as SDF actors) can be statically scheduled as SDF 40]. A larger subset can also be quasi-statically scheduled (thus performing static bu er sizing) by means of a mapping to Petri Nets 47]. While the opportunities for system-level optimization o ered by this choice still need to be fully explored, we can already envision a design ow in which the designer uses multiple languages, depending on the domain of application and other requirements (e.g., tool availability, company policy or personal preference), that all have a semantics in terms of CFSMs. Then multiple scheduling, allocation, partitioning, hardware and software synthesis algorithms can be applied on the CFSM network, possibly depending on the identi cation of special cases that admit an especially e cient implementation. Formal veri cation and simulation can be used throughout the design process, thanks to the re nement-based design applied to a formal model. Re nement occurs both at the functional level, implementing CFSMs ( 6]), and at the communication level, implementing communication ( 56]) and scheduling ( 6, 47]). This scheme has been adopted in the POLIS system and has been also followed by a commercial product of the Alta Group of Cadence Design Systems, Inc. opment of the Tagged Signal Model that has been used throughout this paper, and for his many contributions to the eld of system design for many years. Part of this paper has been adapted from an earlier version co-authored with Prof. Lee and Dr. Stephen Edwards. We wish to thank the Polis team and the Felix team of Cadence Design Systems, Inc., who shared the discoveries and developments which lead to this work. In particular, we wish to thank Dr. Jim Rowson for the key role he has played in putting together the design methodology and the architecture in the Cadence work, and Dr. Felice Balarin 52

Acknowledgments We thank Prof. Ed Lee for the work that led to the devel-

for introducing the notion of equivalence classes between CFSM behaviors. We are indebted to our colleagues of the University of California at Berkeley, of the Politecnico di Torino, of Cadence Berkeley and European Labs and of PARADES for the many discussions and their contributions. This work has been partially sponsored by grants of CNR and Cadence Design Systems.

References

1] W. B. Ackerman. Data ow languages. Computer, 15(2), 1982. 2] R. Alur and D. Dill. Automata for Modeling Real-Time Systems. In Automata, Languages and Programming: 17th Annual Colloquium, volume 443 of Lecture Notes in Computer Science, pages 322{335, 1990. Warwick University, July 16-20. 3] Arvind and K. P. Gostelow. The U-Interpreter. Computer, 15(2), 1982. 4] F. Balarin, H. Hsieh, A. Jurecska, L. Lavagno, and A. SangiovanniVincentelli. Formal veri cation of embedded systems based on CFSM networks. In Proceedings of the Design Automation Conference, 1996. 5] F. Balarin and A. Sangiovanni-Vincentelli. Schedule validation for embedded reactive real-time systems. In Proceedings of the Design Automation Conference, 1997. 6] F. Balarin, E. Sentovich, M. Chiodo, P. Giusto, H. Hsieh, B. Tabbara, A. Jurecska, L. Lavagno, C. Passerone, K. Suzuki, and A. SangiovanniVincentelli. Hardware-Software Co-design of Embedded Systems { The POLIS approach. Kluwer Academic Publishers, 1997. 7] A. Balluchi, M. Di Benedetto, C. Pinello, C. Rossi, and A. SangiovanniVincentelli. Hybrid control for automotive engine management: The cut-o case. In First International Workshop on Hybrid Systems: Computation and Control. LNCS - Springer-Verlag, April 1997. 8] A. Benveniste and G. Berry. The synchronous approach to reactive and real-time systems. Proceedings of the IEEE, 79(9):1270{1282, 1991. 9] A. Benveniste and P. Le Guernic. Hybrid dynamical systems theory and the SIGNAL language. IEEE Transactions on Automatic Control, 35(5):525{ 546, May 1990. 10] G. Berry. A hardware implementation of pure Esterel. In Proceedings of the International Workshop on Formal Methods in VLSI Design, January 1991. 53

11] G. Berry. The foundations of esterel. In G. Plotkin, C. Stirling, and M. Tofte, editors, Proof, Language and Interaction: Essays in Honour of Robin Milner. MIT Press, 1998. 12] S. S. Bhattacharyya, P. K. Murthy, and E. A. Lee. Software Synthesis from Data ow Graphs. Kluwer Academic Press, Norwood, Mass, 1996. 13] G. Bilsen, M. Engels, R. Lauwereins, and J. A. Peperstraete. Static scheduling of multi-rate and cyclo-static DSP applications. In Proc. 1994 Workshop on VLSI Signal Processing. IEEE Press, 1994. 14] F. Boussinot and R. De Simone. The ESTEREL language. Proceedings of the IEEE, 79(9), 1991. 15] R. Bryant. Graph-based algorithms for boolean function manipulation. IEEE Transactions on Computers, C-35(8):677{691, August 1986. 16] J.A. Brzozowski and C-J. Seger. Advances in Asynchronous Circuit Theory { Part I: Gate and Unbounded Inertial Delay Models. Bulletin of the European Association of Theoretical Computer Science, October 1990. 17] J. Buck, S. Ha, E.A. Lee, and D.G. Messerschmitt. Ptolemy: a framework for simulating and prototyping heterogeneous systems. Interntional Journal of Computer Simulation, special issue on SimulationSoftware Development, January 1990. 18] J. T. Buck. Scheduling Dynamic Data ow Graphs with Bounded Memory Using the Token Flow Model. PhD thesis, U.C. Berkeley, 1993. UCB/ERL Memo M93/69. 19] J. R. Burch. Automatic Symbolic Veri cation of Real-Time Concurrent Systems. PhD thesis, Carnegie Mellon University, August 1992. 20] N. Carriero and D. Gelernter. Linda in context. Comm. of the ACM, 32(4):444{458, April 1989. 21] M. Chiodo, P. Giusto, H. Hsieh, A. Jurecska, L. Lavagno, and A. Sangiovanni-Vincentelli. Hardware/software codesign of embedded systems. IEEE Micro, 14(4):26{36, August 1994. 22] F. Commoner and A. W. Holt. Marked directed graphs. Journal of Computer and System Sciences, 5:511{523, 1971. 23] J. Cortadella, M. Kishinevsky, L. Lavagno, and A. Yakovlev. Deriving Petri nets from nite transition systems. IEEE Transactions on Computers, 47(8):859{882, August 1998. 24] J. Desel and J. Esparza. Free choice Petri nets. Cambridge University Press, New York, 1995. 54

25] J. Desel, K.-P. Neuendorf, and M.-D. Radola. Proving non-reachability by modulo-invariants. Theorectical Computer Science, 153(1-2):49{64, 1996. 26] D. Drusinski and D. Harel. On the power of bounded concurrency. I. Finite automata. Journal of the Association for Computing Machinery, 41(3):517{ 539, May 1994. 27] P. Godefroid. Using partial orders to improve automatic veri cation methods. In E.M Clarke and R.P. Kurshan, editors, Proceedings of the Computer Aided Veri cation Workshop, 1990. DIMACS Series in Discrete Mathematica and Theoretical Computer Science, 1991, pages 321-340. 28] N. Halbwachs. Synchronous Programming of Reactive Systems. Kluwer Academic Publishers, 1993. 29] N. Halbwachs, P. Caspi, P. Raymond, and D. Pilaud. The synchronous data ow programming language LUSTRE. Proceedings of the IEEE, 79(9):1305{1319, 1991. 30] D. Harel. Statecharts: A visual formalism for complex systems. Sci. Comput. Program., 8:231{274, 1987. 31] D. Harel, H. Lachover, A. Naamad, A. Pnueli, M. Politi, R. Sherman, A. Shtull-Trauring, and M. Trakhtenbrot. Statemate: A working environment for the development of complex reactive systems. IEEE Trans. on Software Engineering, 16(4), April 1990. 32] T.A. Henzinger. The theory of hybrid automata. Technical Report UCB/ERL M96/28, University of California, Berkeley, 1996. 33] C. A. R. Hoare. Communicating sequential processes. Communications of the ACM, 21(8), 1978. 34] R. Jagannathan. Parallel execution of GLU programs. In 2nd International Workshop on Data ow Computing, Hamilton Island, Queensland, Australia, May 1992. 35] G. Kahn. The semantics of a simple language for parallel programming. In Proc. of the IFIP Congress 74. North-Holland Publishing Co., 1974. 36] D. J. Kaplan et al. Processing graph method speci cation version 1.0. The Naval Research Laboratory, Washington D.C., December 1987. 37] R. M. Karp and R. E. Miller. Properties of a model for parallel computations: Determinacy, termination, queueing. SIAM Journal, 14:1390{1411, November 1966. 38] R. P. Kurshan. Automata-Theoretic Veri cation of Coordinating Processes. Princeton University Press, 1994. 55

39] R. Lauwereins, P. Wauters, M. Ade, and J. A. Peperstraete. Geometric parallelism and cyclostatic data ow in GRAPE-II. In Proc. 5th Int. Workshop on Rapid System Prototyping, Grenoble, France, June 1994. 40] E. A. Lee and D. G. Messerschmitt. Synchronous data ow. IEEE Proceedings, September 1987. 41] E. A. Lee and T. M. Parks. Data ow process networks. Proceedings of the IEEE, May 1995. 42] E. A. Lee and A. Sangiovanni-Vincentelli. The tagged signal model - a preliminary version of a denotational framework for comparing models of computation. Technical report, Electronics Research Laboratory, University of California, Berkeley, CA 94720, May 1996. 43] E.A. Lee and A. Sangiovanni-Vincentelli. Comparing models of computation. In Proceedings of the International Conference on Computer-Aided Design, pages 234{241, 1996. 44] F. Maraninchi. The Argos language: Graphical representation of automata and description of reactive systems. In Proc. of the IEEE Workshop on Visual Languages, Kobe, Japan, October 1991. 45] K. McMillan. Symbolic model checking. Kluwer Academic, 1993. 46] R. Milner. Communication and Concurrency. Prentice-Hall, Englewood Cli s, NJ, 1989. 47] M.Sgroi, L.Lavagno, and A.Sangiovanni-Vincentelli. Quasi-static scheduling of free-choice petri nets. Technical report, UC Berkeley. http://www-cad.eecs.berkeley.edu/~sgroi/qss.ps, 1997. 48] M. Nielsen, G. Rozenberg, and P.S. Thiagarajan. Elementary transition systems. Theoretical Computer Science, 96:3{33, 1992. 49] Institute of Electrical and Electronics Engineers. IEEE standard VHDL language reference manual. IEEE, 1994. 50] J. L. Peterson. Petri Net Theory and the Modeling of Systems. PrenticeHall Inc., Englewood Cli s, NJ, 1981. 51] C. A. Petri. Kommunikation mit Automaten. PhD thesis, Bonn, Institut fur Instrumentelle Mathematik, 1962. (technical report Schriften des IIM Nr. 3). 52] J. Rasure and C. S. Williams. An integrated visual language and software development environment. Journal of Visual Languages and Computing, 2:217{246, 1991. 56

53] W. Reisig. Petri Nets: An Introduction. Springer-Verlag, 1985. 54] F. Rocheteau and N. Halbwachs. Implementing reactive programs on circuits: A hardware implementation of LUSTRE. In Real-Time, Theory in Practice, REX Workshop Proceedings, volume 600 of LNCS, pages 195{208, Mook, Netherlands, June 1992. Springer-Verlag. 55] T. G. Rokicki. Representing and Modeling Digital Circuits. PhD thesis, Stanford University, 1993. 56] J. Rowson and A. Sangiovanni-Vincentelli. Interface-based design. In Proceedings of the Design Automation Conference, pages 178{183, 1997. 57] E.M. Sentovich, K.J. Singh, C. Moon, H. Savoj, R.K. Brayton, and A. Sangiovanni-Vincentelli. Sequential Circuit Design Using Synthesis and Optimization. In Proc of the ICCD, pages 328{333, October 1992. 58] T. R. Shiple, G. Berry, and H. Touati. Constructive analysis of cyclic circuits. In Proceedings of the European Design and Test Conference, March 1996. 59] J. E. Stoy. Denotational Semantics: The Scott-Strachey Approach to Programming Language Theory. The MIT Press, Cambridge, MA, 1977. 60] P. A. Suhler, J. Biswas, K. M. Korner, and J. C. Browne. Td : A task-level data ow language. J. on Parallel and Distributed Systems, 9(2), June 1990. 61] W. Takach and A. Wolf. An automaton model for scheduling constraints in synchronous machines. IEEE Tr. on Computers, 44(1):1{12, January 1995. 62] D.E. Thomas and P. Moorby. The Verilog Hardware Description Language. Kluwer Academic Publishers, 1991. 63] A. Valmari. A stubborn attack on state explosion. Formal Methods in System Design, 1(4):297{322, 1992. 64] P.H.J. van Eijk, C. A. Vissers, and M. Diaz, editors. The Formal description technique Lotos. North-Holland, 1989. 65] M. von der Beeck. A comparison of statecharts variants. In Proc. of Formal Techniques in Real Time and Fault Tolerant Systems, volume 863 of LNCS, pages 128{148. Springer-Verlag, 1994.

57

- A Comparative Evaluation of Models and Specification Languages for Embedded System Design
- 2 Embedded System Design Formal Models for Embedded System Design
- Models of embedded computation
- Design of Embedded Database System
- The design of a clock synchronization sub-system for parallel embedded systems
- The design of a rapid prototype platform for ARM based embedded system
- Facility location models for distribution system design
- A codesign backend approach for embedded system design
- Embedded Control System Design for Autonomous Navigation Mobile Robot
- A framework for embedded system specification under different models of computation in Syst
- Design and Implementation of Scalable and Efficient Programming Models for Fast Computation and Data
- 0_introduce of embedded system
- (译文)Introduction to Embedded Systems
- C语言编程总结
- 电子产品可靠性设计分析方法

更多相关文章：
**Artificial immune algorithm ***for* IIR filter *design*

*embedded* into gradient-descent algorithms have been...*models* proposed *for* NN and evolutionary *computation*...two ?lters were *designed* *for* the *system* identi?...**专业英语**

or a high-end 64-bit microprocessor could provide this*computation* power....the latest techniques *for* *embedded* *system* *design* in a variety *of* applications...**计算机国际核心期刊**

*Design* Automation *of* *Embedded* *Systems* Journal *of* Knowledge-Based Intelligent ... Unconventional *Models* *of* *Computation* MCU: Universal Machines and *Computations* ...**用MATLAB的时重要MATLAB tools**

AIRES - automatic integration*of* reusable *embedded* ...LSDPTOOL - robust control *system* *design* using the...Matrix *Computation* - constructing test matrices, ...**计算机领域内刊物排名**

*Computation* Machine Learning Intl Jnl *of* Computer ... Graphical *Models* & Image Processing CVGIP: Image...*Design* Automation *of* *Embedded* *Systems* Journal *of* ...**matlab 英文资料**

symbolic*computation* and graphics processing power ... *embedded* *system* development, and fixed-point ...*design* and simulation *of* *model* parameters, and so...**国外期刊汇总**

*Computation* (previously known as LISP and Symbolic...*Modeling* and Technology Journal *of* Computers and ...*Design* Automation *of* *Embedded* *Systems* Journal *of* ...**自动化专业英语讲义_学生用**

In many cases, the*design* *of* control *system is*...*of* dealing with more realistic *models* *of* *systems*....tires and the road, as an *embedded* computer ...**国际普遍公认的计算机学科期刊分级**

Graphical*Models* & Image Processing CVGIP: Image...*Design* Automation *of* *Embedded* *Systems* Journal *of* ...*of* *Computation* Mathematical Programming Optimization: ...**个人文档**

*Embedded* *system* *design*,Grid and cloud computing, ...Complex *Systems*: *Modeling* and Simulation ? Hybrid ...Applications *of* *Computation* as a Scientific Paradigm... 更多相关标签：

or a high-end 64-bit microprocessor could provide this

AIRES - automatic integration

symbolic

In many cases, the

Graphical