W. Gropp, E. Lusk, A taxonomy of programming models for symmetric multiprocessors and SMP clusters, in: Programming Models for Massively Parallel Computers, 1995, pp. 2--7.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....5 presents the exported API and Section 6 two representative examples of use of our runtime system. Finally, Section 7 discusses briefly the evolution of our system and concludes. 2. Background 2.1. Programming Clusters of Multiprocessors Clusters of Multiprocessors mix two memory models [7]: each node is an n way SMP and communication between nodes uses message passing hardware. According to the established taxonomy and classification, a CLUMP can use a single memory model (SMM) or a hybrid memory model (HMM) In the SMM approach, a mechanism is provided to unify the memory model ....

W. Gropp and E. Lusk. A taxonomy of Programming Models for Symmetric Multiprocessors and SMP Clusters. In Proceedings of Programming Models for Massively Parallel Computers, pages 2 7, Berlin, Germany, 1995.

....and MPI like message passing paradigms and the simple programming approach (see Fig. 1) Note that our overall algorithmic style is similar in spirit to the one advocated by the bulk synchronous parallel (BSP) model [32] Most popular programming methodologies for Cosmos fall into two categories [15]. The first, distributed shared memory (DSM) systems (for example, TreadMarks [2] from Rice University, Multigrain Shared Memory (MGS) 34] from MIT, and Coherent Virtual Machine (CVM) 19] from University of Maryland) provides a software layer which simulates coherent shared memory between nodes ....

W. W. Gropp and E. L. Lusk, A taxonomy of programming models for symmetric multiprocessors and SMP clusters, in Proceedings of 1995 Programming Models for Massively Parallel Computers, Berlin, Germany, October 1995," pp. 2#7.

....on clusters. We call this approach SIMPLE , referring to the joining of the SMP and MPI like message passing paradigms and the simple programming approach (see Figure 1) SIMPLE is described in more detail in [7] Existing programming methodologies for SMP clusters fall into two categories [12]. The first, distributed shared memory (DSM) systems (for example, TreadMarks [2] from Rice University, Multigrain Shared Memory (MGS) 25] from MIT and Coherent Virtual Machine (CVM) 18] from University of Maryland) provides a software layer which simulates coherent shared memory between nodes ....

W.W. Gropp and E.L. Lusk. A Taxonomy of Programming Models for Symmetric Multiprocessors and SMP Clusters. In Proceedings of 1995.

.... MPI like message passing paradigms and the simple programming approach (see Figure 1) Note that our overall algorithmic style is similar in spirit to the one advocated by the Bulk Synchronous Parallel (BSP) model [32] Most popular programming methodologies for COSMOS fall into two categories [14]. The first, distributed shared memory (DSM) systems (for example, TreadMarks [2] from Rice University, Multigrain Shared Memory (MGS) 34] from MIT and Coherent Virtual Machine (CVM) 19] from University of Maryland) provides a software layer which simulates coherent shared memory between nodes ....

W. Gropp and E. Lusk. A Taxonomy of Programming Models for Symmetric Multiprocessors and SMP Clusters. In Proceedings of 1995.

....results by examining the breakdown of benchmark execution times. Finaly, we compare the performance of our CLUMP with the performance of some high end supercomputers in section 6. 2 Existing approaches to program CLUMP and previous performance evaluations Following the taxonomy and classification [12], 7] there are two basic approaches to program a CLUMP according to the programmer view of memory: a single memory model (SMM) and a hybrid memory model (HMM) In the SMM approach, the programmer only sees one memory model. A mechanism is provided to unify the memory model. Some message passing ....

E.L. Lusk and W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and SMP clusters. In Proceedings of Programming Models for Massively Parallel Computers, pages 2 7, 1995. 9

....are two basic approaches according to the programmer view of memory in a CLUMP: A sitgl roemow model (SMM) and n In SMM approach, the programmer only sees one memory model. A mechanism is provided to unify the memory model. Some message passing libraries have been upgraded to work with CLUMPs [3] [4] [2] The performance results presented in [4] claim that a cluster of uniprocessors can be faster than a CLUMP. Shared virtual memory environments (DSVM) provide another alternative to unify the memory models, as presented in [5] 6] 7] 8] Performance of the DSVM model mainly relies on the ....

....view of memory in a CLUMP: A sitgl roemow model (SMM) and n In SMM approach, the programmer only sees one memory model. A mechanism is provided to unify the memory model. Some message passing libraries have been upgraded to work with CLUMPs [3] 4] 2] The performance results presented in [4] claim that a cluster of uniprocessors can be faster than a CLUMP. Shared virtual memory environments (DSVM) provide another alternative to unify the memory models, as presented in [5] 6] 7] 8] Performance of the DSVM model mainly relies on the efficiency of the memory consistency protocol ....

E. L. Lusk W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and SMP clusters. In in Proceedings of Programming Models for Massively Parallel Computers, pages 2 7, 1995.

....CLUMP and previous perfor mance evaluations Networks of multiprocessors PC present an hybrid memory model: message passing between nodes and shared memory inside each node. Several ways exist to program CLUMPs. Taxonomy and classification of programming model for the CLUMPs have been reviewed [1] [2] We may classify two main approaches by distinguishing the programmer view of the memory in a CLUMP: a single memory model (SMM) an hydride memory model (HMM) In SMM approach, programmer see only one memory model. A mechanism is provided in the platform to unify the view of memory ....

E. L. Lusk W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and smp clusters. In m Proceedings of Programming Models for Massivell Parallel Computers, pages 2 7, 1995.

....les diffdrents modales de programmarion disponibles ou envisageables pour les CLUMPs. 2 Les approches existantes pour programmer les CLUMPs et les r6sultats de performance existants I1 existe diffdrentes approches pour programmer les CLUMPs. Des classifications ont ddj dtd prdsentdes dans [1] et [2] Nous pouvons distinguer principalement deux approches en fonction de la faqon dont le programmeur voit la mdmoire. I1 peut voit: un seul modale mdmoire (SMM) un modale mdmoire hybride (HMM) Dans l approche SMM, un mdcanisme est ajoutd la plate forme afin que le programmeur ne ....

E. L. Lusk W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and smp clusters. In m Proceedings of Programming Models for Massivell Parallel Computers, pages 2 7, 1995.

....use multiprocessor PCs as the nodes for a parallel platform. Networks of multiprocessors PC present an hybrid memory model: message passing between nodes and shared memory inside each node and conform to the term of CLUMP (CLUster of MultiProcessors) It exists several ways to program the CLUMPs. [1] and [2] presents respectively a taxonomy and a classification of the programming model for the CLUMPs. We may classify two main approaches by distinguishing the programmer view of the memory in a CLUMP: a single memory model (SMM) or an hydride memory model (HMM) In the SMM approach the ....

E. L. Lusk W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and smp clusters. In in Proceedings of Programming Models for Massively Parallel Computers, pages 2-7, 1995.

....and HMM) Section 6 helps to understand the speedup of dual processor versions by analyzing breakdowns of the execution time for all benchmarks and by measuring the memory traffic inside a dual processor node. 2 CLUMP programming models According to the established taxonomy and classification [2] [3] a CLUMP can use a single memory model (SMM) or an hybrid memory model (HMM) In the SMM approach, a mechanism is provided to unify the memory model that the programmer sees. Some message passing libraries have been updated to provide CLUMP with a single memory model [4] 1] 3] Using ....

E. L. Lusk W. W. Gropp. A taxonomy of programming models for symmetric multiprocessors and SMP clusters. In in Proceedings of Programming Models for Massively Parallel Computers, pages 2--7, 1995.

.... MPI like message passing paradigms and the simple programming approach (see Figure 1) Note that our overall algorithmic style is similar in spirit to the one advocated by the Bulk Synchronous Parallel (BSP) model [32] Most popular programming methodologies for COSMOS fall into two categories [15]. The first, distributed shared memory (DSM) systems (for example, TreadMarks [2] from Rice University, Multigrain Shared Memory 1 cosmos ( kaz m os) noun Greek kosmos c. 1650 1: an orderly harmonious systematic universe 2: a complex orderly self inclusive system 3: Cluster Of Shared Memory ....

W. W. Gropp and E. L. Lusk. A Taxonomy of Programming Models for Symmetric Multiprocessors and SMP Clusters. In Proceedings of 1995 Programming Models for Massively Parallel Computers, pages 2--7, Berlin, Germany, October 1995.

....of a wide variety of parallel architectures. In the past decade, this variety has gradually reduced to two main alternatives: shared memory MIMD systems of modest size (perhaps 30 processors) using cache coherence, sometimes gathered into larger ensembles that have come to be called clumps [15]; and distributed memory MIMD systems based on commodity processor boards and off the shelf interconnects, called clusters [1] This picture may change with the widespread deployment of optical computing devices or other new technology, but at present almost all deployed systems are of these two ....

W. W. Gropp and E. L. Lusk. A taxonomy of programming models for symmetric multiprocessors and SMP clusters. In Programming Models for Massively Parallel Computers, pages 2--7, October 1995.

....for implementing scientific applications on dual tier computers remains elusive. The programmer must carefully orchestrate parallelism and locality in the application, managing the interaction of processes, threads, shared memory, message passing, synchronization, scheduling, and load balancing [3, 4, 5]. Such software techniques are beyond the reach of many application programmers, and the lack of effective software tools hinder efficient implementations of scientific calculations on dual tier architectures by the scientific community. This paper presents a domain specific programming ....

....can disrupt the execution of tightly optimized loop nests, lowering CPU performance. The techniques required to work around this difficulty, e.g. mixed mode programming, are beyond the means of many programmers. The interaction between message passing and threads is difficult to understand [3, 4, 5]. 4 2.4 Summary We have now identified a set of system and application requirements. In sum, we require run time data decomposition and collective communication models that may be customized to the needs of the application and even to the specific input. Communication is assumed to be expensive ....

W.W. Gropp and E.L. Lusk, "A Taxonomy of Programming Models for Symmetric Multiprocessors and SMP Clusters," Programming Models for Massively Parallel Computers, W.K. Giloi, S.Jahnichen, and B.D. Shriver, eds, IEEE Computer Society Press, 1995, pp. 2--7.

....for implementing scientific applications on dual tier computers remains elusive. The programmer must carefully orchestrate parallelism and locality in the application, managing the interaction of processes, threads, shared memory, message passing, synchronization, scheduling, and load balancing [3, 4, 5]. Such software techniques are beyond the reach of many application programmers, and the lack of effective software tools hinder efficient implementations of scientific calculations on dual tier architectures by the scientific community. This paper presents a domain specific programming ....

....It can also lead to severe performance losses by disrupting tightly optimized loop nests. The techniques required to work around this difficulty, e.g. multi threading, are beyond the means of most programmers. The interaction between multi threading and message passing is difficult to understand [3, 4, 5]. 2.4 Summary We have now identified a set of system and application requirements. In sum, we require run time data decomposition and collective communication models that may be customized to the needs of the application and even to the specific input. Communication is assumed to be expensive and ....

W. W. Gropp and E. L. Lusk, "A taxonomy of programming models for symmetric multiprocessors and SMP clusters," in Proc. 1995: Programming models for massively parallel computers, pp. 2--7, Oct. 1995.

....brings together several pieces of high performance interconnection technology: bus backplanes for symmetric multiprocessors, low latency networks for connections between machines, and simple, user level primitives for communication. Several groups have studied the problem of programming Clumps [3, 6, 11, 13, 15, 16]. Some of these efforts focus on issues related to shared virtual memory [17, 32] but most relate to high level message libraries such as MPI. The software overheads associated with memory allocation and tag matching in traditional message passing libraries often obscure the machine level ....

W. W. Gropp, E. L. Lusk, "A Taxonomy of Programming Models for Symmetric Multiprocessors and SMP clusters," Proceedings of Programming Models for Massively Parallel Computers 1995, October 1995, pp. 2-7.

....a straightforward hierarchical extension of the XYZ levels would cover the processor level. In general, it may be worthwhile to consider a recursive XYZ program model based on either nested CTAs or the PMH. Gropp and Lusk have proposed a programming model taxonomy for SMPs and SMP clusters [73]. Their taxonomy classifies programming styles along three axes: address space, process scheduling, and heterogeneity of the model. It is not clear how a hybrid model like an XYZ program fits in Gropp and Lusk s taxonomy. Crandall et. al [51] report experiences with dual level parallel programs on ....

W. W. Gropp and E. L. Lusk. A taxonomy of programming models for symmetric multiprocessors and SMP clusters. In Proceedings 1995: Programming models for massively parallel computers, pages 2--7, October 1995. 165

....to portable parallel programming. However, their extension to multi tier parallelism remains an unresolved issue. At present, the programmer faces myriad options to manage heavyweight processes, lightweight threads, shared memory, messagepassing, synchronization, scheduling, and load balancing [4]. This daunting array of low level detail hinders efficient multi tier implementation. We present a small set of programming abstractions to simplify implementation of efficient algorithms for block structured scientific calculations on dedicated SMP clusters. This paper extends previous work [5] ....

W. W. Gropp and E. L. Lusk, "A taxonomy of programming models for symmetric multiprocessors and SMP clusters," in Proceedings 1995: Programming models for massively parallel computers, pp. 2--7, October 1995.

....the proper programming model for multi tier parallel computers remains an unresolved issue. At present, the programmer faces myriad options regarding the coordination of heavyweight processes, lightweight threads, shared memory, message passing, synchronization, scheduling, and load balancing [4]. This daunting array of lowlevel programming detail hinders efficient implementations for multi tier platforms. Stephen Fink was supported by the DOE Computational Science Graduate Fellowship Program, Scott Baden by NSF contract ASC 9520372. Computer time on the Maryland Digital AlphaServer ....

W. W. Gropp and E. L. Lusk, "A taxonomy of programming models for symmetric multiprocessors and SMP clusters," in Proceedings 1995: Programming models for massively parallel computers, pp. 2--7, October 1995.

....nodes. 5.2 Implications for Programming Models The proper programming paradigm for SMP clusters is presently an unresolved issue. An SMP cluster application may manage message passing, lightweight threads, and shared memory in a of variety of styles and combinations to achieve high performance [12, 13, 14]. Unfortunately, managing these constructs can drown the application programmer in a morass of low level implementation details, discouraging parallel implementations. A solution to this problem lies in higher level programming models that shield the programmer from low level details of ....

W. W. Gropp and E. L. Lusk, "A taxonomy of programming models for symmetric multiprocessors and SMP clusters," in Proceedings 1995: Programming models for massively parallel computers, pp. 2--7, October 1995.

No context found.

W. Gropp, E. Lusk, A taxonomy of programming models for symmetric multiprocessors and SMP clusters, in: Programming Models for Massively Parallel Computers, 1995, pp. 2--7.

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