Tolerating latency in software distributed shared memory systems through multithreading

grantor: University of TorontoLarge communication latency is a key obstacle to achieving higher performance on software distributed shared memory (DSM) systems, which provide a shared memory abstraction on message passing hardware. Multithreading is one of the techniques that can tolerate the communication latency on software DSM. It tolerates the latency by overlapping communication and computation with multiple threads on each processor. We investigate the performance of multithreading on software DSMs empirically. Our results demonstrate that multithreading can improve the speedups of the applications by 3% to 84%, and we can further improve the performance of some applications by modifying the source codes or making threads switch context on synchronization only. This thesis also compares the performance of multithreading with that of prefetching and investigates their combined effects. The best overall approach to hiding latency depends on the predictability of memory access patterns and the extent to which lock stalls dominate synchronization time.M.Sc