In many parallel applications, network latency causes a dramatic loss in processor utilization. THis paper presents KaHPF, an optimizing HPF compiler, that uses software controlled access pipelining (SCAP) to hide communication latency. Various benchmarks of common, but complex communication patterns demonstrate KaHPF's performance improvement on a Cray T3E. KaHPF's compiled code is faster than code that uses highly optimized shared-memory system functions by a factor of 1.2 to 3.2. In comparison th PGI's HPF compiler KapHPF shows an improvement of 3.9 to 30.4.
展开▼