Dynamic data distributions offer a number of performance benefits, but require more sophisticated compiler support and incur run-time overhead. We investigate attainable benefits using a compiler system we developed for the Hector NUMA multiprocessor. We show that the benefits depend on a number of factors, includingdata size relative to the cache size, data access patterns, the degree of "NUMAness" of the multiprocessor system, and the extent to which data is reused. Programmers and compiler designers must take these factors into consideration.
Tarek S. Abdelrahman, Kenneth L. Ma