| M. A. Nichols, H. J. Siegel, H. G. Dietz, "Data management and control-flow aspects of an SIMD/SPMD parallel language/compiler," IEEE Transactions on Parallel and Distributed Systems, Vol. 4, No. 2, February 1993, pp. 222-234. |
....parallelism; including most massively parallel applications [ Fox, 1988 ] 1 Some of these steps may not be fully specified at compilation time. Some of the responsibility for the parallel mapping to the architecture may be left to the compiler, or the run time system [Anderson et al. 1995, Nichols et al. 1993]. 10 Main Memory Data Instructions Control Unit Datapath instruction stream data stream Figure 2.1: The Von Nuemann Model 3. Degree of parallelism. This property determines the number processing elements an application can successfully utilize to speed up the execution. Applications utilizing ....
....is sequential. There is only one program to write and debug, and the data distribution is often intuitive. In fact, due to the simplicity of the data parallel model used to program SIMD machines, it has become a popular programming model for MIMD machines [ Brandes, 1993, Hatcher and Quinn, 1991, Nichols et al. 1993 ] 4. Memory efficiency. On MIMD organizations, the PE memory must hold both the program and data. Only data are kept in the SIMD PE memory. The saving in memory cost and size is considerable when we consider a massively parallel configuration [ Manning and Meyer, 1993b ] The major ....
[Article contains additional citation context not shown here]
M. A. Nichols, H. J. Siegel, and H. G. Dietz. Data management and control-flow aspects of an SIMD/SPMD parallel language/compiler. IEEE Transactions on Parallel and Distributed Systems, pages 222--234, February 1993.
....with Reconfigurable Busses[6] In order to exploit the advantages of the data parallel programming model without mapping it on the SIMD computational model much work has been done in the direction of mapping it on a different computational model. Both the definition of the SPMD computational model [7] [8] and the new CM 5 system from the Thinking Machines Corp. 9] can be classified in this stream. Unfortunately, in our opinion the first approach has only partially solved the problems previously mentioned in that it proposes enhancements that do not deal with the implementation problems of the ....
.... in our opinion the first approach has only partially solved the problems previously mentioned in that it proposes enhancements that do not deal with the implementation problems of the clock distribution and the Table 1 Features Summarize SIMD (e.g. MP 1 [4] SBMD SPMD (e.g. PASM [7]) Data Parallel MIMD (e.g. CM 5 [9] Programming Model Implicitly Synchronous Implicitly Synchronous Usually Explicitly Synchronized Explicitly Synchronized Program Control Flow Global Local Local Local CPU Architecture Very Simple. Many on a single Chip. Simple. Many on a single Chip Usually ....
M. A. Nichols, H. J. Siegel and H. G. Dietz, Data Management and Control Flow Aspects of an SIMD/SPMD Parallel Language/Compiler, IEEE Trans. on Parallel and Distributed Systems, Vol. 4, No. 2, pp. 222-234, February 1993.
....Busses[8] In order to exploit the advantages of the data parallel programming model without mapping it on the SIMD computational model a second approach has been to go in the direction of mapping it on a different computational model. Both the definition of the SPMD computational model [9], the CM 5 system from the Thinking Machines Corp. 5] can be classified in this stream. Finally, following a third approach, other scientists have proposed to introduce a limited degree of relaxation of the SIMD paradigm synchronization in such a way as to keep it invisible to the machine users ....
.... is interesting to observe that the Data Flow computational paradigm has gone through a similar process evolving from early single instruction Data Flow machines like the Manchester Data Flow or Sigma 1, to macro block Data Flow machines like EM 4 [11th] SIMD (e.g. MP 1) SBMD SPMD (e.g. PASM [9]) Data parallel MIMD Programming Model Implicitly Synchronous Implicitly Synchronous Usually Explicitly Synchronized Explicitly Synchronized Program Control Flow Global Local Local Local CPU Architecture Very Simple. Many on a single Chip. Simple. Many on a single Chip Usually off the shelf CPUs ....
M. A. Nichols, H. J. Siegel and H. G. Dietz, Data Management and Control Flow Aspects of an SIMD/SPMD Parallel Language/Compiler, IEEE Trans. on Parallel and Distributed Systems, Vol. 4, No. 2, pp. 222-234, February 1993.
....Connection Machine language C [23] calls these variables scalar. There is no equivalent of our inference system for these languages, as the properties we are inferring are guaranteed by SIMD semantics. Our proposed single keyword provides similar advantages for SPMD languages. The ELP language [21] [24] a joint SIMD SPMD programming language where both modes have the same semantics, allows declaration of single valued variables with a mono keyword. When in SPMD mode the compiler guarantees that the single valued property is preserved, presumably using rules similar to ours (the paper ....
M. A. Nichols, H. J. Siegel, and H. G. Dietz. Data Management and Control-Flow Aspects of an SIMD/SPMD Parallel Language/Compiler. IEEE Transactions on Parallel and Distributed Systems, 4(2):222--234, February 1993.
....radiosity 11319 5 no, not SPMD raytrace 10020 1 yes water 1744 9 yes 2971 9 (both versions) volrend 3704 13 yes cholesky 5050 4 yes fft 1005 7 yes lu 988 5 yes 763 5 (both versions) radix 879 7 yes Table 2: Results of examining the SPLASH 2 benchmarks. SPMD languages. The ELP language [21, 24], a joint SIMD SPMD programming language where both modes have the same semantics, allows declaration of single valued variables with a mono keyword. When in SPMD mode the compiler guarantees that the single valued property is preserved, presumably using rules similar to ours (the paper does ....
M. A. Nichols, H. J. Siegel, and H. G. Dietz. Data Management and Control-Flow Aspects of an SIMD/SPMD Parallel Language/Compiler. IEEE Transactions on Parallel and Distributed Systems, 4(2):222--234, February 1993.
....In addition, accessibility is not necessarily correlated with locality. Split C exposes the locality by providing special operations (put and get) to access remote memory [36, 37] A number of systems go a step further and introduce special annotations that tag variables as private or shared [48, 49]. This is useful as an indication of where the variable should be stored in a NUMA architecture. For example, on the Cedar system the degree of sharing dictates whether a variable is stored in a cluster memory or in the global memory [50, 51] An alternative approach is to initially allow all ....
Mark A. Nichols, Howard Jay Siegel, and Henry G. Dietz. Data management and control-flow aspects of an SIMD/SPMD parallel language/compiler. IEEE Trans. Parallel & Distributed Syst., 4(2):222--234, Feb 1993.
....where each processor sends a ready signal to a control processor which, in turn, informs all processors of arrival at the barrier. The upper bound of 6 Cn assumes p = 64 processors and a log p Cn overhead. 6 Percentage improvement for Obar = 3Cn . Pandore [APT90] Kali [MR90] EPL [NSD93] Parlation Lisp [Sab87] and pC [LG93] Dataparallel languages use the universally addressable memory model and typically provide optional data distribution primitives to specify how the program data is distributed over the memory hierarchy of the parallel architecture. Although some languages ....
....statement, FORALL, and the INDEPENDENT directive. The latter asserts that the statements in the body of the subsequent DO or FORALL statement do not exhibit any sequentializing dependencies, allowing statement instances to be executed independently. Explicit Language for Parallelism (EPL) NSD93] is targeted for the PASM [SSK 81] a partitionable SIMD MIMD parallel processing system. EPL allows the mode of execution (SIMD or SPMD) to be specified explicitly at an arbitrary level of granularity. EPL, however, requires that the user maintain coherency between copies of data present on ....
M.A. Nichols, H.J. Seigel, and H.G. Dietz. Data Management and Control-Flow Aspects of an SIMD/SPMD Parallel Language/Compiler. IEEE Transactions on Parallel and Distributed Systems, 4(2):222--234, February 1993.
....in this thesis. CHAPTER 3 ELP SIMULATOR 3.1. Introduction In this chapter, the design and implementation of a simulator for a mixed mode parallel language targeted to the PASM prototype (see Subsection 2.4.2) is discussed. The E ## xplicit L ## anguage for P ## arallelism (ELP) [NiS93, SiB96, Wan94] is a parallel language designed for programming mixed mode parallel machines; that requires programmers to explicitly specify the mode of parallelism to be used to execute each instruction. ELP provides constructs for both SIMD and MIMD parallelism and an ELP application program is able to ....
M. A. Nichols, H. J. Siegel, H. G. Dietz, "Data management and control-flow aspects of an SIMD/SPMD parallel language/compiler," IEEE Transactions on Parallel and Distributed Systems, Vol. 4, No. 2, February 1993, pp. 222-234.
....The appendix reviews the basic probability theory and notation used here. 2 Overview of the Approach Applications for this study are assumed to be data parallel programs written in a modeindependent language for execution on an SIMD SPMD mixed mode machine. A modeindependent language (e.g. see [NiS93, WeW94]) is a language in which syntactic elements have interpretations under more than one mode of parallelism, and operations represent the most explicit level at which the program representation is identical for each mode of parallelism. Such languages make it possible to utilize the most appropriate ....
M. A. Nichols, H. J. Siegel, and H. G. Dietz, "Data Management and Control-Flow Aspects of an SIMD/SPMD Parallel Language/Compiler," IEEE Transactions on Parallel and Distributed Systems, Vol. 4, No. 2, pp. 222--234, Feb. 1993.
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