V. P. Srini, "An Architectural Comparison of Dataflow Systems," Computer, Vol. 19, No. 3, Mar. 1986. pp. 68-88,  Home/Search   Document Not in Database   Summary   Related Articles   Check

....in dynamic architectures, nodes can be created at run time (e.g. to support loop unravelling and recursion) Second, in static architectures, at most one instance of an actor may be enabled for firing at a time. Dynamic architectures support several instances of an actor firing simultaneously[1]. Finally, static architectures use the same storage space for instructions (actors) and data (tokens) i.e. impure code) Dynamic architectures use physically separate memories for instructions (i.e. pure code) and data[2] In static dataflow architectures, the dataflow program graph is ....

V. P. Srini, "An Architectural Comparison of Dataflow Systems," Computer, Vol. 19, No. 3, Mar. 1986. pp. 68-88,

....actor firings. This scheduling can be done at compile time or at run time, and in the latter case, can be done by hardware or by software. The most widely known execution models for dataflow process networks have emerged from research into computer architectures for executing dataflow graphs [5][93]. This association may be unfortunate, since the performance of such architectures has yet to prove competitive [49] In such architectures, actors are fine grained, and scheduling is done by hardware. Although there have been some attempts to apply these architectures to signal processing [77] ....

V. Srini, "An Architectural Comparison of Dataflow Systems," Computer, 19(3), March 1986.

.... programming language [Grimshaw93a] Grimshaw93b] The control mechanism that selects for execution the primitive units of computation, i.e. sequential operations, is based in CAP on the macro dataflow MDF model [Grimshaw93a] inspired by the dataflow computational model [Agerwala82] Denning86][Srini86][Veen86] The CAP macro dataflow computational model is a coarse grain, data driven model and differs from traditional dataflow in four ways. First, the computation granularity is larger than in traditional dataflow. The basic units of computation are high level tasks such as multiplying two ....

V. P. Srini, "An Architectural Comparison of Dataflow Systems," IEEE Computer, March 1986, Vol. 19, No. 3, 68-88

....waiting for it. Instead of fetching the same data several times, copies of data are produced and simultaneously sent to the instructions waiting for it. In the past 15 years quite a few projects have embarked on the design for a dataflow machine. For a survey of design published before 1986 see [14]. Many of these early projects did not get past the design stage and only two were bold enough to announce that their goal was to produce a commercial dataflow machine: the static dataflow project at Hughes Aircraft Co. 19] and the Data Driven Signal Processor Project at ESL Inc. 5] Due to ....

V.Srini. An Architectural Comparison of Dataflow Systems, IEEE Computer 19(3), 68-88, 1986.

....Also dynamic and irregular applications might cause excessive waste of cycles when mapped to a blocking thread model. 142 8. 2 Software Multithreaded Systems In the classical strict data flow model of computation, an instruction is enabled for execution when all its operands are available [66, 85, 63, 65, 68, 47, 155, 70, 77, 127, 130, 86, 123, 125, 12, 97, 133, 132, 124, 15, 17, 45, 57, 150, 140]. To enforce the enabling condition, the instructions that produce such operands must be able to send a synchronization signal to all the instructions that will consume the recently produced result. This model proved unyielding for the implementation of machines based on current standard ....

Vason P. Srini. An architectural comparison of dataflow systems. Computer, 19(3):68--88, Mar. 1986.

....aspects and performance of two of the run time system implementations, on a network of Sun SparcStation 2 workstations and an Intel Paragon. 2. Background 2. 1 Macro Dataflow The macro dataflow (MDF [19] model is a medium grain, data driven computation model inspired by dataflow [2] 16][41][47] Recall that in dataflow, programs are directed graphs where the vertices are computational primitives (e.g. add, subtract, compare, etc. called actors, the edges, or arcs, model data dependencies, and tokens carry data along the edges between the actors. An actor is a function that maps ....

....were straight line code, more complicated fragments, such as those that contain loops, conditionals, and multiple scopes, are handled in the same fashion. 3.2.3. Token Matching In the macro dataflow model, as in pure dataflow, tokens belonging to a particular computation must be matched [41][47] For example, in Figure 6(a) the tokens containing the values 4 and 5 must be matched. When both tokens are available, and they have been matched, we say the actor is enabled and may fire (execute) The matching process is complicated by the fact that there may be more than one instance of ....

[Article contains additional citation context not shown here]

V. P. Srini, "An Architectural Comparison of Dataflow Systems," IEEE Computer, pp. 68-88, March, 1986.

....the network arcs and used as input for computation at subsequent FUs. Since the dataflow approach [11] is asynchronous, FUs may perform their computations in different speeds. Even though theories exist about dataflow models [4] 14] 22] 25] and many architectures have been proposed (see [19] [28], and [30] for an insightful survey) they can be grouped into static and dynamic (or tagged token) models. Recently a new proposal, which has the advantages of both models, has been presented in [1] The static model allows only one token at a time to reside on an arc, while the dynamic one ....

Srini V.P., "An Architectural Comparison of Data Flow Systems", IEEE Computer, March 1986.

....a parallel algorithm as being composed of a very large number of fine grain parallel activities. Much research has been done on the mapping and scheduling of such fine grain activities so as to achieve high performance. Dataflow architectures use extensive hardware support to achieve this goal [9, 25]. The common approach on conventional multiprocessors is to increase the granularity at compile time based on an analysis of the program structure (see, e.g. 22] We advocate the support of ultra light weight tasks in the runtime system, and contend that efficient support is possible at run ....

V. P. Srini, "An architectural comparison of dataflow systems". Computer 19(3), pp. 68--88, Mar 1986.

....have a distinct right solution. Moreover, some data flow machines include special execution units that handle data structures more efficiently. Various machines that fall in the category of data flow computers have been built and simulated. Comparative reviews of data flow machines are given in [25, 26, 30]. In the following we discuss how our results can be related to existing data flow machines. Machine Architecture The first issue that affects the applicability of our results is the communication structure of the machine. Our models fits some of the possible architectures as described below. We ....

V.P. Srini. An architectural comparison of dataflow systems. Computer, 19(9):68--88, March 1986.

....function invocation and is unaware that the function is implemented in parallel. 2.1. The Macro Dataflow Model of Computation Macro dataflow is Mentat s underlying model of computation. The macro dataflow (MDF [11] model is a medium grain, data driven computation model inspired by dataflow[1] 9][24][28] Recall that in dataflow, programs are directed graphs where the vertices are computational primitives (e.g. add, subtract, compare, etc. called actors, the edges, or arcs, model data dependencies, and tokens carry data along the edges between the actors. Dataflow is data driven in that ....

....RV is actual. IF the RV is delayed, it constructs an arg struct that points to the computation instance that will generate the value. F 16 3.2.3. Token Matching and Predicate Support In the macro data flow model, as in pure data flow, tokens belonging to a particular computation must be matched [24][28] For example, in Figure 10 (a) the tokens containing the values 4 and 5 must be matched. When both tokens are available, and they have been matched, we say the actor is enabled and may fire (execute) The matching process is complicated by the fact that there may be more than one instance of ....

[Article contains additional citation context not shown here]

V. P. Srini, "An Architectural Comparison of Dataflow Systems," IEEE Computer, pp. 68-88, March, 1986.

....of the overhead involved, such operations would only be issued where operand availability (in time or space) could not be assured. Finegrain data flow and task flow is similar in that operations are triggered by data availability, but is inefficient since that is the only execution mechanism [19,27]. Exo ops are part of procedural code and so the programmer does not have to cast the application s control flow into an unfamiliar data flow paradigm. Hybrid or large grain data flow improves efficiency by executing more of the code as conventional processors would. One such proposed system is ....

V. Srini, "An architectural comparison of dataflow systems," IEEE Computer, vol. 19, no. 3, pp. 6888, March 1986.

....actors and arcs transmitting tokens which carry the values to be processed by the actors. Actors become activated for execution when tokens are present at their input arcs. Even though theories exist about dataflow models [3] 6] 10] 11] and many architectures have been proposed (see [8][13][14] for an insightful survey) they can be grouped into static and dynamic (or tagged token) models. Recently a new proposal, which gains the advantages of both models, has been presented in [1] The static model allows only one token at a time to reside on an arc, while the dynamic one allows ....

V.P. Srini, "An Architectural Comparison of Data Flow Systems", IEEE Computer, March 1986.

....and termination) A dynamic partitioning strategy must find some compromise between the benefit of added concurrency and the drawback of added overhead. Some have avoided this problem to some extent by relying on specialized hardware to reduce the cost of managing tasks. Dataflow machines [ Srini, 1986, Arvind and Nikhil, 1990 ] and multithreaded architectures [ Halstead and Fujita, 1988, Nikhil et al. 1991, Agarwal, 1991 ] fall in this category. However, software methods are attractive because they offer portability and low hardware cost. This thesis explores software methods for lowering the ....

V. P. Srini. An architectural comparison of dataflow systems. IEEE Computer, 19(3):68--88, March 1986.

....at subsequent FUs . Since the dataflow approach [DEN80] is asynchronous, FUs may perform their computations with different delays. Even though theories exist about dataflow models [BOH83] DEN85] KAV86] MIK84] LEE95] and many architectures have been proposed (see [GAU91] LEEB94] [SRI86], and [VEE86] for an insightful survey) the problem of how to map directly dataflow graph programs onto hardware is still an open issue [LEEB94] One of the reasons is the mismatch between dataflow models and dataflow machines. The fined grain approach to parallelism must be reviewed in terms of ....

Srini V.P., "An Architectural Comparison of Data Flow Systems", IEEE Computer, March 1986.

No context found.

Srini, V. P., "An Architectural Comparison of Data-Flow Systems," IEEE Computer, pp. 68-88, March, 1986.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC