作者： D. Brent Weatherly , Donald G. Morris , Franklin Lowenthal , David K. Lowenthal
DOI:
关键词: Parallel computing 、 Key (cryptography) 、 Computer science 、 Software distributed shared memory 、 Initial distribution 、 Compiler 、 Time system 、 Distributed computing 、 Distribution (mathematics) 、 Context (language use) 、 Dynamic data
摘要: Distributing data is one of the key problems in implementing efficient distributed-memory parallel programs. The problem especially difficult programs where (1) redistribution between computational phases considered or (2) participating processors (nodes) executing a application are not dedicated. In either case, commonly usedBLOCK andCYCLIC distributions no longer suffice. We have investigated this context software distributed shared memory (SDSM) system. developed integrated compile- and run-time analysis for SDSM systems to support both single- multi-phase applications on potentially non-dedicated workstation clusters. Our system, SUIF-Adapt, selects an effective initial distribution per phase, finds global distribution, adapts changes underlying computing environment, including dynamic removal addition nodes. Each these features contributes good performance; particular, SUIF-Adapt performs significantly better than hand-coded static many distributions.