抄録
Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (A), Fund Type: -, Overall Grant Amount: - (direct: 26900000, indirect: 8070000)
We have studied the OpenMP programming environment for the next generation 64-bit high-performance clusters, by using software distributed shared memory (SDSM) system to enable OpenMP program to run on the cluster. We have also developed a programming support system for OpenMP, and numerical libraries using OpenMP.
1.We ported the SCore cluster system software to 64-bit processor architectures. We conducted the performance evaluation of SCASH DSM system which runs on SCore.
2.We have designed and implemented a very portable SDSM system, SCASH-MPI which uses MPI as its communication layer. MPI is the most portable communication library supported for many kinds of high-speed communication network, so that this approach provide highly portability It allows the users to make use of wide address space in 64-bit processor. We found that the overhead of this implementation is just 6% comparing to the original SCASH.
3.We have designed a new SDSM system, FDSM, by using the access pattern analysis of applications. The access pattern is detected by a hardware mechanism provided by IA64, and is used for efficient communication. It achieves more performance than SCASH.
4.We have studied the optimization of OpenMP program running a DSM system of heterogeneous clusters. We found that the performance can be improved by the combination of the loop re-partitioning and the page migration.
5.We have designed and implemented the interactive tool, OMP/iPat, to support the programmer for OpenMP program developments. It allows the programmer to develop his OpenMP program interactively using the information from parallelism analysis by the compiler.
6.We have conducted the performance evaluation by using the OpenMP benchmark, SPEC-OMP. We have designed and implemented an algorithm of parallel recursive FFT by using OpenMP for IA-64 shared memory multi-processors.