High Performance Computing and Grid in Asia Pacific Region, International Conference on (1997)
Apr. 28, 1997 to May 2, 1997
Jae Bum Lee , Seoul National University,KOREA
Chu Shik Jhon , Seoul National University,KOREA
Cache memories are widely accepted in shared-memory multiprocessor system because they make it possible to reduce network traffics and memory latencies. However, they impose substantial overheads of cache coherency maintenance and also engender some inefficiencies of coherence misses.This paper presents that constraints of cache coherency can be relaxed in a region where exclusive accesses are guaranteed by synchronization operations, so that the execution time of parallel programs can be decreased. When an acquire operation occurs, cache coherence protocol could not be applied to the operations following an acquire operation until a release operation is issued. The updated data in the region are transferred to other processors when a release operation is performed. We alleviate the overheads of release operations using QOLB synchronization primitives. By a program-driven simulation, our new cache coherence protocol shows performance improvement in most of parallel applications, and also execution times can be reduced effectively as the number of processors are increased.
J. B. Lee and C. S. Jhon, "Relaxing Cache Coherence Protocol with QOLB Synchronizations," High Performance Computing and Grid in Asia Pacific Region, International Conference on(HPCASIA), Seoul, Korea, 1997, pp. 1.