
896
CHAPTER
11 Latency Tolerance
dynamically scheduled processors with nonblocking reads, relaxed models are help-
ful in reducing read stall time as well as write stall time. Prefetching also helps
reduce both, so it is interesting to examine whether starting from sequential consis-
tency performance is helped more by prefetching only or by using only a relaxed
consistency model. Even when all optimizations to improve the performance of
sequential consistency are applied (like hardware prefetching, speculative reads, and
write buffering), it is found to be more advantageous to use a relaxed model without
software prefetching than to use a sequentiall ...