A.L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474--482, April 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of such programs differ, many have highly regular sharing behaviors. The set of shared data accessed by individual threads is often invariant from one iteration to the next. This regular behavior can be used by DSMs to predict future accesses, and to move data in advance of subsequent accesses [4, 11, 12]. Such update protocols allow much of the latency of remote data fetches to be hidden. Given reasonably efficient communication, DSMs should be able to achieve good speedups on such applications. The output of parallelizing compilers, such as SUIF [13] is a good source for this type of ....

....memory environments requires far less analysis. Further, the set of applications that can currently be 2 analyzed well enough to turn into a shared memory application is much larger than for message passing applications. By combining parallelizing technology with sophisticated runtime systems [3, 4, 12], we can create a programming environment that is flexible and easy to use. Scientists are not required to write message passing programs or use data parallel languages such as HPF. Instead, they can write sequential programs, rewriting a few computation intensive procedures, and adding ....

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel, "Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers," in Proceedings of the International Parallel Processing Symposium, 1997.

....than its one dimensional counterpart. Unfortunately, two dimensional distributions are not eciently supported in software distributed shared memory (DSM) systems. Such systems provide an attractive shared memory programming model on a distributed machine [LH89, CBZ91, KDCZ94, Ift98, LCD 97, CDLZ97, KT97] In a software DSM, two dimensional distributions can cause a large amount of excess communication This work was supported by National Science Foundation CAREER Grants CCR 9733063 and CCR 9876073. y Department of Computer Science, The University of Georgia, Athens, GA 30602. ....

Alan L. Cox, Sandhya Dwarkadus, Honghui Lu, and Willy Zwanapoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474-482, April 1997.

....of such programs differ, many have highly regular sharing behaviors. The set of shared data accessed by individual threads is often invariant from one iteration to the next. This regular behavior can be used by DSMs to predict future accesses, and to move data in advance of subsequent accesses [1 3]. Such update protocols allow much of the latency of remote data fetches to be hidden. Given reasonably efficient communication, DSMs should be able to achieve good speedups on such applications. The output of parallelizing compilers, such as SUIF [4] is a good source for this type of ....

....memory environments requires far less analysis. Further, the set of applications that can currently be analyzed well enough to turn into a shared memory application is much larger than for message passing applications. By combining parallelizing technology with sophisticated runtime systems [1, 3, 5], we can create a programming environment that is flexible and easy to use. Scientists are not required to write message passing programs or use data parallel languages such as HPF. Instead, they can write sequential programs, rewriting a few computation intensive procedures, and adding ....

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel, "Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers, " in Proceedings of the International Parallel Processing Symposium, 1997.

....gather nonlocal data. A second approach is to combine shared memory compilers (e.g. SUIF [11] with software DSM systems (e.g. TreadMarks [30] CVM [24] which provide a shared memory interface. Software DSMs are less efficient than explicit messages, but are much simpler compilation targets [4,25]. In this paper, we examine existing approaches to parallelizing irregular reductions, and propose a new efficient algorithm. 1.1 Irregular Reductions We begin by looking at the example irregular reduction shown in Figure 1. The computation loops over the edges of an irregular graph, computes a ....

....is inherently less efficient than sending explicit messages. Compilers for software DSMs have adopted a number of techniques for improving efficiency. One approach relies on precise communication analysis to insert explicit messages when analysis exactly identifies interprocessor communication [3,4,7]. A second approach exploits customized coherence protocols for reductions and nonlocal updates [8,24,25,33,39] In addition, compilers can also eliminate unnecessary synchronization based on communication analysis [17,25] TreadMarks is one of the most efficient software DSMs currently ....

[Article contains additional citation context not shown here]

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, April 1997.

....second approach is to combine shared memory compilers (e.g. SUIF [6] with software distributed sharedmemory (DSM) systems (e.g. TreadMarks [17] CVM [13] which provide a shared memory interface. Software DSMs are less efficient than explicit messages, but are much simpler compilation targets [3, 14]. In this paper, we introduce LocalWrite, a new compiler and run time parallelization technique which can improve performance for certain classes of irregular reductions. We evaluate the performance of different parallelization approaches as we vary application characteristics, in order to ....

....is inherently less efficient than sending explicit messages. Compilers for software DSMs have adopted a number of techniques for improving efficiency. One approach relies on precise communication analysis to insert explicit messages when analysis exactly identifies interprocessor commu nication [2, 3, 5]. A second approach exploits customized coherence protocols for reductions and nonlocal updates [13, 14, 18] In addition, compilers can also eliminate unnecessary synchronization based on communication analysis [7, 14] TreadMarks is one of the most efficient software DSMs currently available. It ....

[Article contains additional citation context not shown here]

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, April 1997.

....can be used to improve data movement at inefficient points in application executions. Our future work with tapes will center on two areas. First, we are exploring the use of compilers to automatically generate tapes interfaces. This work is complementary to recent work in parallelizing compilers [8, 32]. Tapes improve performance by exploiting repetitive access patterns. Identifying such patterns with high degree of probability in the compiler is much easier than generating explicit messagepassing code for the data movement. Hence, compiler heuristics that might not be rigorous enough to ....

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel, "Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers," in Proceedings of the International Parallel Processing Symposium, 1997.

....applications on software DSM systems shows nearly the same as messagepassing systems, but with simpler compiler support. Several researches also investigated combining software DSMs and explicit message passing aided by compilers in order to enhance general performance of software DSM systems [8, 16, 11]. This hybrid approach improves the performance of regular computations as well as irregular computations, making software DSMs more attractive on message passing machines. Without compiler support, software DSMs also exploit iterative nature of scientific applications by prefetching the same ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, Apr. 1997.

....the node s memory by sending a message to the node that has the page. Hence, internode communication is done implicitly by the DSM system in a manner similar to how disk I O is managed in paged operating systems. Recently, DSM systems have received attention as attractive compiler targets[KT97, CDLZ97] This is because it is much easier to generate code without having to determine at compile time whether data is local or not. Furthermore, when communication patterns cannot be determined at compile time, compilers often have to generate code with all to all communication, which is prohibitively ....

Alan L. Cox, Sandhya Dwarkadus, Honghui Lu, and Willy Zwanapoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474--482, April 1997.

....in DSM systems. However, their work primarily concerns local (single phase or two adjacent phases) data distribution. Finally, many have studied integrated compiler DSM systems with a focus on elimination of as many consistency actions as possible using compiler information. KT97, LCD 97, CDLZ97, MHS94] 6 Summary We have described the design and implementation of an integrated compiler run time system for global data distribution in distributed shared memory (DSM) systems. The SUIF Adapt system efficiently supports a larger class of applications than previous compiler only approaches. ....

Alan L. Cox, Sandhya Dwarkadus, Honghui Lu, and Willy Zwanapoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474--482, April 1997.

....programs are portable since they can be run on the large scale parallel machines as well as the low end, but more pervasive multiprocessor workstations. Shared memory parallelizing compilers are easy to use, flexible, and can accept a wide range of applications. Results from several recent studies [4, 14] indicate they can approach the performance of current message passing compilers or explicitly parallel message passing programs on distributed memory machines. However, load imbalance and synchronization overhead were identified as sources of inefficiency when compared with message passing ....

....they were able to compute data access patterns at compile time and use it to help the runtime system aggregate communication and synchronization. Cox et al. conducted an experimental study to evaluate the performance of TreadMarks as a target for the Forge SPF shared memory compiler from APR [4]. Results show that SPF TreadMarks is slightly less efficient for densematrix programs, but outperforms compiler generated message passing code for irregular programs. They also identify opportunities for the compiler to eliminate unneeded barrier synchronization and aggregating messages in the ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, April 1997.

....shared memory parallelizing compilers with software distributed shared memory (DSM) systems that provide a coherent shared address space in software. Shared memory parallelizing compilers are easy to use, flexible, and can accept a wide range of applications. Results from several recent studies [1, 5] indicate they can approach the performance of current messagepassing compilers or explicitly parallel message passing programs on distributedmemory machines. Unfortunately, load imbalance and synchronization overhead were identified as sources of inefficiency in compiler parallelized programs. ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, April 1997.

....and gather nonlocal data. A second approach is to combine shared memory compilers (e.g. SUIF [7] with software DSM systems (e.g. TreadMarks [19] CVM [15] which provide ashared memory interface. Software DSMs are less efficient than explicit messages, but are much simpler compilation targets [3, 16]. In this paper, we introduce LOCALWRITE, a new compiler and run time parallelization technique which can improve performance for certain classes of irregular reductions. We evaluate the performance of different parallelization approaches as we vary application characteristics, in order to ....

....but is inherently less efficient than sending explicit messages. Compilers for software DSMs have adopted a number of techniques for improving efficiency. One approach relies on precise communication analysis to insert explicit messages when analysis exactly identifies interprocessor communication [2, 3, 5]. A second approach exploits customized coherence protocols for reductions and nonlocal updates [6, 15, 16, 20, 25] In addition, compilers can also eliminate unnecessary synchronization based on communication analysis [9, 16] TreadMarks is one of the most efficient software DSMs currently ....

[Article contains additional citation context not shown here]

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, Apr. 1997.

....from IBM and DEC. 0 20 40 60 80 100 barrier imbalance seq wait comm OS application Figure 1. Breakdown of Total Execution Time (16 Processor SP 2) Shared memory parallelizing compilers are easy to use, flexible, and can accept a wide range of applications. Results from recent studies [3, 14] indicate they can approach the performance of current message passing compilers or explicitly parallel messagepassing programs on distributed memory machines. However, load imbalance and synchronization overhead were identified as sources of inefficiency when compared with message passing ....

....they were able to compute data access patterns at compile time and use it to help the runtime system aggregate communication and synchronization. Cox et al. conducted an experimental study to evaluate the performance of TreadMarks as a target for the Forge SPF shared memory compiler from APR [3]. Results show that SPF TreadMarks is slightly less efficient for dense matrix programs, but outperforms compiler generated message passing code for irregular programs. They also identify opportunities for the compiler to eliminate unneeded barrier synchronization and aggregating messages in the ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, Apr. 1997.

.... tolerance [12] In particular, even if written assuming a shared memory programming model, a program in which communication patterns are amenable to compile time analysis might make better use of compiler generated message passing communication than of even hardware supported shared memory [14, 12]. Third, distributed shared memory and message passing implementations are naturally similar at a low level, so if an implementation provides shared memory in hardware the additional cost of exposing message passing is low [40] Finally, one reading of technology trends is that shared memory ....

Alan L. Cox, Sandhya Dwarkadas, Honghui Lu, and Willy Zwaenepoel. Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers. In Proceedings of the 11th International Symposium on Paralel Processing, April 1997.

....second approach is to combine shared memory compilers (e.g. SUIF [8] with software distributed shared memory (DSM) systems (e.g. TreadMarks [24] CVM [18] which provide a shared memory interface. Software DSMs are less efficient than explicit messages, but are much simpler compilation targets [4, 20]. In this paper, we introduce LOCALWRITE, a new compiler and run time parallelization technique which can improve performance for certain classes of irregular reductions. We evaluate and compare the performance of different approaches to supporting irregular computations as a function of how ....

....but is inherently less efficient than sending explicit messages. Compilers for software DSMs have adopted a number of techniques for improving efficiency. One approach relies on precise communication analysis to insert explicit messages when analysis exactly identifies interprocessor communication [3, 4, 6]. For irregular computations, the compiler can prefetch index array sections to reduce latency [24] A second approach exploits customized coherence protocols for reductions and nonlocal updates [7, 20, 25, 30] Less precise communication analysis is required, since the compiler does not need to ....

[Article contains additional citation context not shown here]

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, April 1997.

....compiler achieved a linear speedup for NAS EP (slightly modified) but no speedup at all for NAS FT from the NPB 2. x suite running on an 8 processor Alpha SMP[5] Parts of the NPB1 has been implemented for distributed shared memory (ThreadMarks) and message passing (PVM) as described in [10] and [3]. On a cluster of 8 HP735 connected with 100 Mb s FDDI network they achieved speedups of 7.9 for NPB EP and 4.41 5.47 for NAS FT[10] On an 8 processor IBM SP 2 they achieved of speedups of 3.06 5.12 for NPB FT[3] These works mention only small data sets for NAS FT (size S and A) Some previous ....

....shared memory (ThreadMarks) and message passing (PVM) as described in [10] and [3] On a cluster of 8 HP735 connected with 100 Mb s FDDI network they achieved speedups of 7.9 for NPB EP and 4.41 5.47 for NAS FT[10] On an 8 processor IBM SP 2 they achieved of speedups of 3.06 5. 12 for NPB FT[3]. These works mention only small data sets for NAS FT (size S and A) Some previous work using Dolphin SCI interfaces for real applications exist. Performance of a SAR (Synthetic Aperture Radar) application on workstations clustered with the previous generation of SCI interfaces is measured in ....

[Article contains additional citation context not shown here]

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallizing compilers. In Proceedings of Eleventh International Parallel Processing Symposium, Apr. 1997.

....dominated by the barrier time in most applications. When compared to message passing programs, additional sources of overhead for traditional software DSM systems have been identified to include separation of data and synchronization, overhead in detecting memory faults, and absence of aggregation [6]. Researchers have attempted to use compiler assisted analysis of the program to reduce these overheads. Prefetching of pages has been suggested by a number of groups for improving the performance of TreadMarks, by saving the overhead of a memory fault [19, 17] This technique sacrifices the ....

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers. In Proceedings of International Parallel Processing Symposium, April 1997.

....performance on architectures with much longer communication latencies than the Myrinet interconnect, a more difficult task. Concurrent with our work, Cox et al. conducted an experimental study to evaluate the performance of TreadMarks as a target for the Forge SPF shared memory compiler from APR [4]. They compared its performance against the message passingcode generated by the Forge xHPFcompiler,as well as hand coded shared memory and message passing versions of the program. Results show that SPF TreadMarks is slightly less efficient for dense matrix programs, but outperforms ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, Geneva, Switzerland, Apr. 1997.

.... are examined in the context of a prototype system [17] extending SUIF [11] to target the CVM [15] software distributed shared memory (DSM) Recent studies indicate such systems can approach the performance of current message passing HPF compilers and explicit messagepassing MPI programs [6, 17]. However, load imbalance and synchronization overhead were identified as key sources of inefficiency. Execution times for five SUIF parallelized applications on a 16 processor IBM SP 2 are shown in Figure 1. Total execution time is split into application processing time, OS overhead, ....

.... to improve their performance on a software DSM [7] Chandra and Larus evaluated combining the PGI HPF compiler and the Tempest software DSM system [4] Cox et al. conducted an experimental study to evaluate the performance of TreadMarks as a target for the Forge SPF shared memory compiler from APR [6]. They identify opportunities to eliminate unneeded barrier synchronization; many of their suggestions are implemented in the SUIF CVM system. Rajamony and Cox developed a performance debugger for detecting unnecessary synchronization at run time by instrumenting all loads and stores [21] In the ....

A. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel ProcessingSymposium, Geneva, Switzerland, April 1997.

....34.26 MRI 1135.06 3448.16 3.00 3.12 SOR 1537.91 1480.63 16.09 17.38 TSP 588.33 947.28 7.45 9. 36 Table 2 Communication characteristics (per node) ware DSM systems have been identified to include separation of data and synchronization, overhead in detecting memory faults, and absence of aggregation [21]. Researchers have attempted to use compiler assisted analysis of the program to reduce these overheads. Prefetching of pages has been suggested by a number of groups for improving the performance of TreadMarks, by saving the overhead of a memory fault [17,8] This technique sacrifices the ....

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel, "Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers, " in Proceedings of International Parallel Processing Symposium, April 1997.

No context found.

A.L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474--482, April 1997.

No context found.

A.L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474--482, April 1997.

....a uniform and perhaps a more intuitive programming paradigm. A shared memory paradigm provides ease of use and additionally leverages an SMP workstation s available hardware coherence to handle sharing within the SMP. SDSM has been shown to be an e ective target for a parallelizing compiler [5]. Since data caching and communication is implemented by the run time system, compile time complexity is reduced. Previous work [7, 19] has integrated compile time information within the run time system in order to improve performance. Access patterns from the compiler are used by the run time ....

A.L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474-482, April 1997.

....whether uniprocessors or symmetric multiprocessors (SMPs) offer cost effective and highly available parallel computing environments. Software distributed shared memory (SDSM) provides a shared memory abstraction on a distributed memory machine, with the advantage of ease ofuse. Previous work [5] has shown that an SDSM run time can prove to be an effective target for a parallelizing compiler. The advantages of using an SDSM system include reduced complexity at compile time, and the ability to combine compile time and run time information to achieve better performance ( 6, 18] One issue ....

A.L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel. Evaluating the performance of software distributed shared memory as a target for parallelizing compilers. In Proceedings of the 11th International Parallel Processing Symposium, pages 474-- 482, April 1997.

No context found.

A. L. Cox, S. Dwarkadas, H. Lu, and W. Zwaenepoel, "Evaluating the Performance of Software Distributed Shared Memory as a Target for Parallelizing Compilers," in IPPS, April 1997. 10

First 50 documents

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