Frederic T. Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Technical Report MIT/LCS/TR-697, MIT Laboratory for Computer Science, August 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....in Table 4.1. In each case, the producer consumer relationship for each communication is known at compile time, so explicit producer initiated mechanisms can be used. Many numerical applications have well defined communication patterns that are known at compile time or after a pre processing phase [18]. We do not consider benchmarks with dynamic or migratory communication patterns. Implicit mechanisms, such as invalidate, update or selective update, may be better suited to those environments. Explicit mechanisms can be used, however, if the consumer can be predicted or if the cost of ....

....are also of interest. These are representative of a large class of applications in engineering, for instance. One of the challenges of this type of application is to map the communication mesh efficiently onto the parallel system, maximizing parallelism while minimizing communication. Chong [18] and others have done excellent work in this area, and we hope that more of these codes will be used as benchmarks in future studies. Larger systems and benchmarks. While one of the motives for parallel systems is the ability to solve problems faster, another is the ability to solve much larger ....

Frederic Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Technical Report MIT/LCS/TR-697, Laboratory for Computer Science, MIT, October 1996.

.... 18000 particles, 6 iterations Table 1: Applications and Kernels MP3D, Barnes, Locus, Chol, and Water are from the SPLASH suite [SWG92] Appbt and MG are part of the NAS parallel benchmarks [Bai94] The rest of the applications are engineeringtype kernels from the University of Rochester, MIT [CA96] and Berkeley [CDG 93] We categorize the applications into traditional, coarse grain applications and emerging, fine grain applications. The traditional, coarse grain applications are applications that appear in most studies of sharedmemory applications and architectures. These include ....

Frederic T. Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Technical Report MIT/LCS/TR-697, MIT Laboratory for Computer Science, August 1996.

....obtained from the Harwell Boeing benchmark suite [12] We started from existing parallel iccg algorithms [20] 41] and implemented an interrupt message passing version. The remaining four versions were all derived from the initial messagepassing code. Implementation details are available in [9]. 4.3.1 iccg with Message Passing The iccg computation graph is essentially a dataflow computation [3] and is easily implemented via active messages. Non local edges, i.e. edges between nodes on different processors, are communicated with active messages. Each processor keeps a presence counter ....

Frederic T. Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Technical report, mitlcs -tr-697, MIT Laboratory for Computer Science, Cambridge, MA, October 1996.

.... the ICCG sparse triangular solve kernel running on a large structural finite element matrix, the BCSSTK32 2 million element automobile chassis, obtained from the Harwell Boeing benchmark suite [9] We started from existing parallel ICCG algorithms [14] 26] Implementation details are available in [7]. ICCG with Message Passing The ICCG computation graph is essentially a dataflow computation [2] and is easily implemented via active messages. For each node in its local memory, a processor keeps track of when all incoming edges have been satisfied, whereafter the outgoing edges can be ....

Frederic T. Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Tech rpt, mit-lcs-tr-697, MIT Lab for Comp Sci, Cambridge, MA, 1996.

....of each of the applications and their input parameters. MP3D, Barnes, Locus, Chol, and Water are from the SPLASH suite [SWG92] Appbt and MG are part of the NAS parallel benchmarks [Bai94] The rest of the applications are engineeringtype kernels from the University of Rochester, MIT [CA96] and Berkeley [CDG 93] We categorize the applications into traditional, coarse grain applications and emerging, fine grain applications. The traditional, coarse grain applications are applications that appear in most studies of shared2 Coarse Grain Description Input Data Program Appbt ....

Frederic T. Chong and Anant Agarwal. Shared memory versus message passing for iterative solution of sparse, irregular problems. Technical Report MIT/LCS/TR-697, MIT Laboratory for Computer Science, August 1996.

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