G. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. PhD thesis, Massachusetts Institute of Technology,1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....each equation could be a separate thread enabled for execution when the data it requires becomes available [14, 1, 4] Such synchronization would enforce a valid order of EPL program execution. However, this form of parallelization is not efficient on current dataflow machines, such as the Monsoon[5], because all synchronization is done at run time, increasing the overhead incurred by token communication and matching. A more efficient approach is to recognize at compile time which threads must execute sequentially relative to each other and then to merge them. This is the motivation behind ....

G. M. Papadopoulos. Implementation of a general-purpose dataflow multiprocessor. Research Monographs in Parallel and Distributed Computing. MIT Press, 1991.

....frequencies. Smith advocates a shift from instruction level parallelism to instruction level distributed processing with more emphasis on inter instruction communication along with dynamic optimization and a tight interaction between hardware and low level software [3] Dataflow architecture [4, 5, 6] is an alternative to the von Neumann model. However previous attempts to develop practical systems based on dataflow model have failed for numerous reasons [15] Hybrid models that combine the two alternatives have also been explored. Our SDF architecture can be viewed as a new hybrid approach. ....

. Papadopoulos, G. M., "Implementation of a general purpose dataflow multiprocessor," Tech. Report TR432, Laboratory for computer Science, MIT, Cambridge, MA, Aug. 1988.

....of memory where the locations are determined as a function not only of the destination address but also of the context. This way, two left tokens waiting for their partners at the same node will be stored in different locations. This format is modeled after Papadopoulos explicit token store (ETS) [17]. Sample code for a plus node in the dynamic system is: A1 NSG.CONTEXT 4This approach would be less efficient if the encoding of the instruction to load a constant into RO on the J Machine were not so simple. To load the constant X into RO, one simply places X in the instruction stream. ....

Papadopoulos, Gregory Michael. Implementation of a General Purpose Dataflow Multi- processor. Technical Report MIT/LCS/TR-432, MIT Laboratory for Computer Science, Cambridge, MA, 1989. (PhD Thesis, Department of EECS, MIT.) 106

....are called finely multithreaded processors, and the others are called coarsely multithreaded processors or block multithreaded processors. Several processor designs have used multithreading to mask communication and synchronization latencies, or to utilize deep pipelines effectively, e.g. [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. By multithreading a processor such that an instruction from a different thread can be initiated every cycle (or every few cycles) pipeline bubbles due to pipeline dependencies or processor stalls due to memory latency can be prevented. Processors in message passing multicomputers often maintain ....

....to mitigate these problems, such as special frames for pipeline state, might adversely impact the processor cycle time. The opposing goals of high single thread performance and fast context switches have been previously addressed largely in their extremes. Finely multithreaded processors [2, 3, 5, 7] that disallow the execution of consecutive instructions from the same process can support very fast context switches, because the various instructions in the pipeline at any given time are independent. Consequently, they can use multithreading to utilize deep pipelines efficiently, in addition to ....

G.M. Papadopoulos. Implementation of a GeneraLPurpose Data flow Multiprocessor. Tech- nical Report TR-432, MIT Laboratory for Computer Science, Cambridge, MA, August 1988.

....(i.e. remote operations, like memory reads and procedure calls) and their consumers. Tokens are used not only to signal that data values are ready, but also to preserve the original program order be tween instructions which have side effects. Most of the previous work on data flow machines [24, 31, 14] could dispense with this feature because it was handling functional languages. For instance, two store instructions that have no data dependency between them cannot be reordered if they may update the same memory location. 4 Resources Required for ASH Implementations In this Section we present ....

Gregory Michael Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. Technical Report MIT/LCS/TR-432, Laboratory for Computer Science, Massachusetts Institute of Technology, 1988.

....general purpose mechanisms for communication, synchronization, and naming that will support a wide variety of parallel models of computation. We have developed two experimental dataflow programming systems for the J Machine. For the first system, we adapted Papadopoulos explicit token store [10] to implement static and then dynamic dataflow. Our second system made use of Iannucci s hybrid execution model [8] to combine several dataflow graph nodes into a single sequence, decreasing scheduling overhead. The conclusion describes how the strengths of the two systems can be combined. We have ....

....at its destination, it is placed on the message queue, and a new task is created when the message reaches the head of the queue. Overview In the next section, we describe a straightforward method of implementing the dataflow language Id [9] based on Papadopoulos explicit token store (ETS) [10]. In the following section, we describe the system we built to simulate Iannucci s hybrid architecture [8] on the J Machine. In the conclusion, we discuss the strengths and weaknesses of the two systems and describe our plans to combine them into an efficient implementation of Id. ETS Static ....

[Article contains additional citation context not shown here]

Gregory Michael Papadopoulos, Implementation of a General Purpose Dataflow Multiprocessor, MIT Laboratory for Computer Science Technical Report 432, Cambridge, MA, 1989. (PhD Thesis, Department of EECS, MIT.)

....of memory where the locations are determined as a function not only of the destination address but also of the context. This way, two left tokens waiting for their partners at the same node will be stored in different locations. This format is modeled after Papadopoulos explicit token store (ETS) [17]. Sample code for a plus node in the dynamic system is: A1 MSG.CONTEXT 4 This approach would be less efficient if the encoding of the instruction to load a constant into R0 on the J Machine were not so simple. To load the constant X into R0, one simply places X in the instruction stream. ....

Papadopoulos, Gregory Michael. Implementation of a General Purpose Dataflow Multiprocessor. Technical Report MIT/LCS/TR-432, MIT Laboratory for Computer Science, Cambridge, MA, 1989. (PhD Thesis, Department of EECS, MIT.) 106

....and temporally unpredictable often implemented using hashing. ffl Uncontrolled fan out can course a token explosion problem. 31 6 Anaconda a real time control flow data flow hybrid processor Tagged token dynamic data flow An alternative paradigm (used for the MIT Monsoon machine [10]) is to remove operand storage from the instruction and to place it in a data page, or activation frame . Activation frames effectively store the context of a function. This may be implemented using a conventional flat memory store with the addition of an empty full flag for each word in memory. ....

.... operations fetching their operands from the register file (i.e. the control flow segments are limited to at most two input parameters) The problems with tagged token data flow are: ffl Every time a function is instantiated an activation frame has to be emptied (128 words on the Monsoon machine [10]) ffl Matching the first operand for a dyadic operation results in no operation to be performed and thus a bubble in the pipeline. Conclusions about current data flow paradigms Static data flow, though simple, is impractical because functions cannot be supported properly. Coloured data flow ....

G.M. Papadopoulos. Implementation of a general-purpose dataflow multiprocessor. Research Monographs in Parallel and Distributed Computing. MIT Press, 1991.

....of a module or a group of modules. Another system component is the Module Builder, which is used to create modules without doing any programming beyond that needed to implement the module s function. 3. Data flow architectures have been investigated from a hardware perspective as well [Gur85, Pap91] The execution model offers attractive properties for parallel processing implicit synchronization of parallel activities and self schedulability. Unlike the von Neumann model which explicitly states the sequence of instructions, in the data flow model the execution of any instruction is ....

....and self schedulability. Unlike the von Neumann model which explicitly states the sequence of instructions, in the data flow model the execution of any instruction is driven by the operand availability. This emphasizes a high degree of parallelism at the instruction level. The Monsoon Project [Pap91] developed by MIT and Motorola produced a data flow multiprocessor targeted to large scale scientific and symbolic computation. Its success motivated much of the work on similar projects [NPA92] and contributed to spread the interest in data flow and parallel programming. 4. Avionics Control ....

[Article contains additional citation context not shown here]

Gregory M. Papadopoulos. Implementation of a general-purpose dataflow multiprocessor. MIT Press, 1991.

....the thread firing rule (which determines when threads are enabled) can be based on either a blocking or a non blocking strategy. The blocking strategy is adopted in Iannucci s Hybrid Architecture [18] the Tera MTA [2] and the EARTH machine [17] The non blocking strategy is adopted in Monsoon [26, 27], T [23] and the EM 4 [31] among others.The Threaded Abstract Machine (TAM) 11] is a software implemented multithreaded execution that has been ported to a number of platforms (such as the TMC CM 5 and the Cray T3D) it implements the non blocking model. In this section we present an ....

G. Papadopoulos. Implementation of a general-purpose dataflow multiprocessor. Technical report TR-432, MIT Laboratory for Computer Science, August 1988.

....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 ....

Gregory Michael Papadopoulos. Implementation of a general purpose dataflow multiprocessor. Tech. Rep. MIT/LCS/TR-432, MIT Lab. for Comp. Sci., Aug. 1988. PhD thesis.

....software support for SPMD execution to execute these data parallel operations. At this time, it is unclear how this can be accomplished. This is a matter for further investigation. 5.2.5. Dataflow Translation Several dataflow architectures are currently being constructed at various universities [20, 21]. Although considerably less developed than the architectures already discussed, dataflow architectures are of interest to a significant number of researchers in parallel computing. Our example dataflow architecture is the Monsoon [20] The Monsoon machine consists of a number of processing ....

....are currently being constructed at various universities [20, 21] Although considerably less developed than the architectures already discussed, dataflow architectures are of interest to a significant number of researchers in parallel computing. Our example dataflow architecture is the Monsoon [20]. The Monsoon machine consists of a number of processing elements with a small amount of local memory, connected to memory elements by a multistage packet switching network. This system contains some special hardware for supporting dataflow graph execution. The computation nodes of a VMPP program ....

Gregory M. Popadopoulos, "Implementation of a General Purpose Dataflow Multiprocessor ", Technical Report TR432, MIT Lab for Computer Science, Cambridge, MA, September, 1988.

....of multithreaded hybrid dataflow von Neumann architectures have been proposed and built. They lie on different points along the line between dataflow and von Neumann models. They range from the Tera [ACC 90] and the T [NPA92, Bec92] which are close to the von Neumann model, to the Monsoon [Pap88] and the EM 4 EM 5 [SYH 89, SYH 91] which are close to the dataflow model. 1.4 Code Generation Strategies for Multithreaded Architectures While generating code for multithreaded machines, various factors have to be considered. These include balancing the resource utilization, load ....

....models introduced are TAM, P RISC and the Spineless Tagless G machine. Section 2.3 deals with code generation methods for the multithreaded machines and models. 2.1 Multithreaded Architectures 2.1. 1 MIT Monsoon The Monsoon is a realization of the Explicit Token Store (ETS) architecture [Pap88] ETS provides a simple model of dynamic dataflow execution. The ETS architecture itself is an outgrowth of the MIT Tagged Token Dataflow Architecture (TTDA) ABM87] The properties of the TTDA that the designers of the ETS wanted to retain were: ffl large synchronization name space, ffl ....

G. M. Papadopoulos. Implementation of a General-Purpose Dataflow Multiprocessor. Technical Report TR-432, MIT Laboratory for Computer Science, August 1988.

....significant parallelism are the rule level parallelism and the LHS level parallelism. The latter reduces the matching time and the former reduces the total number of recognize act cycles. 2. 3 Data Flow Computations Most data flow machines such as MIT Tagged Token Dataflow Architecture[1] Monsoon[23, 24], EM 4[26] were designed for general purpose parallel numerical computations. As for data flow symbolic computations, GAMMA[6] AGM[2] KARDAMOM[4] have demonstrated that the data flow principle can be applied in database applications. We will demonstrate that data flow principle can also be ....

Gregory M. Papadopoulos. Implementation of a General-Purpose Dataflow Multiprocessor. MIT Press, 1991.

....designers of parallel machines have typically redesigned the entire processor, incorporating aggressive message processing into the architecture and into the programming model. A good example of this strategy is provided by the Monsoon multiprocessor developed at MIT in the late 1980 s [Pap91] Monsoon can perform interprocessor communication with no performance overhead. Monsoon s strategy for interprocessor communication is to incorporate the network into the processor pipeline. Figure 1 4 outlines the Monsoon processor pipeline. As in a conventional RISC processor, each processor ....

....the processor. Again, designs which use this approach do not implement the general message passing model. Examples of these designs include shared memory machines, such as the shared memory interface of the MIT Alewife machine [ACD 91] and dataflow machines, such as the MIT Monsoon machine [Pap91] Since the messages are controlled without software intervention, they can be handled very efficiently. For example, a Monsoon processor can receive, dispatch on, and create messages at the rate of one per cycle. However neither of the two provides explicit user level model of the network. The ....

Gregory M. Papadopoulos. Implementation of a general-purpose dataflow multiprocessor. The MIT Press, Cambridge, MA, 1991.

....machines [6, 11, 12, 17] As described at a recent performance tools workshop [5] however, many issues in this area are still open. In this paper, we describe a performance tool that combines data collection, analysis and visualization into an integrated environment. It was developed for Monsoon[14, 3, 18], a multithreaded multiprocessor machine built by Motorola in collaboration with MIT. The tool was designed to be both easy to use and complete, in that it provides almost all the information useful for evaluating parallel execution. This paper describes our tool in detail with examples ....

G. M. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. PhD thesis, Massachusetts Institute of Technology, Cambridge MA, August 1988.

....Emulation Facility, but it was never implemented. Sigma 1 [22] from Japan s Electro Technical Laboratory (ETL) was designed and built based on TTDA, and the Manchester Dataflow Machine [20] was also implemented based on similar dynamic dataflow concepts. Recent hardware projects include Monsoon [30], EM 4 from ETL [24] and Epsilon 2 from Sandia National Labs [19] We will use the abstract TTDA machine as a basis to explain the operation of a typical pipelined dataflow processor. Figure 2.2 provides a conceptual view of the TTDA pipeline. A datum or token coming from the Token Queue enters ....

....the complexity of implementing hashing functions in hardware, even the best schemes require three pipeline cycles to execute [30, p.45] 2.2. 1 Monsoon The need for a better implementation of the Waiting Matching section was the primary motivation behind Papadopoulos s Monsoon Dataflow Processor [30]. 3 Monsoon is actually an implementation of the more general Explicit Token Store (ETS) architecture. The basic idea behind the ETS architecture is to do away with the expensive Waiting Matching stage of the TTDA pipeline by having each token tag encode the address of a unique memory 3 This ....

Gregory M. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. PhD thesis, MIT, August 1988.

....the Monsoon project. For 2 A thread is simply a sequential fragment of code, where each instruction executes after the one before it. sequential programming, Monsoon instruction set may be viewed as a single accumulator type of instruction set (reminiscent of the earliest von Neumann machines) [30]. We did not need any experimental evidence to show that Monsoon cannot compete with a modern RISC processor in executing Fortran or C 1.1 Overview In Section 2, we discuss our execution model for parallel programs, the Monsoon system, including its hardware and software, and our implementation ....

G. M. Papadopoulos. Implementation of a General-Purpose Dataflow Multiprocessor. Research Monograph in Parallel and Distributed Computing. MIT Press, 1992.

....are executed simultaneously within and across processors. Moreover, given sufficient parallelism, dynamic instruction scheduling has the added pragmatic benefit of being resilient to long and unpredictable communication latency. The most recent generation of dataflow machines (e.g. MIT s Monsoon [11, 12], ETL s EM 4 [13] and Sandia s Epsilon 2 [6] have shown how operand matching can be accomplished with simple hardware structures in two machine cycles. There does seem to be an unavoidable price of purely dynamic instruction scheduling, however. Each dyadic (two input) instruction requires the ....

....an I Fetch to an empty location is deferred by saving the return continuation in the location, and a subsequent I Store causes the fetch to be satisfied. A more detailed explanation, including what happens when more than one IFetch request arrives before the corresponding I Store, may be found in [11]. Monsoon also supports a pair of mutual exclusion operations called Take and Put. The format of requests for I Fetch I Store and for Take Put is exactly like that described for Fetch Store; only the processing of these requests at the receiving node differs. 3.4 Continuation Manipulation ....

G. M. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. Technical Report TR432, MIT Laboratory for Computer Science, 545 Technology Square, Cambridge, MA, September 1988. (PhD Thesis, Dept. of EECS, MIT).

No context found.

G. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. PhD thesis, Massachusetts Institute of Technology,1988.

No context found.

G. Papadopoulos. Implementation of a General Purpose Dataflow Multiprocessor. PhD thesis, Massachusetts Institute of Technology, 1988.

No context found.

G.M. Papadopoulos. "Implementation of a general purpose dataflow multiprocessor", Tech Report TR432, Laboratory for Computer Science, MIT, Aug. 1988.

No context found.

Gregory M. Papadopoulos, "Implementation of a General-Purpose Dataflow Multiprocessor", Technical Report TR-432, MIT Lab. for Comp. Sci., Aug., 1988.

No context found.

Papadopoulos, G.M. --- Implementation of a General Purpose Dataflow Multiprocessor --- MIT/LCS Technical Report 432, August 1988 (PhD thesis)

No context found.

G. M. Papadopoulos. Implementation of a General-Purpose Dataflow Multiprocessor. Technical Report TR-432, MIT Laboratory for Computer Science, August 1988.

First 50 documents  Next 50

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