1 Evaluating Dynamic Prefetching in a Software DSM

While software-controlled static prefetching in software DSMs appears to achieve great performance on many applications, the results in investigating prefetching with runtime information are not satisfied due to the formidable overhead. Seeking specific prefetching mechanisms to reduce the execution time of a certain kind of applications, however, is a promising work and will impact on SVMs and other fields. In this report, an experimental work, adding dynamic prefetching in JIAJIA, a software DSM based on scope consistency [Pal 96], has been proposed and evaluated. There are two features in our prefetching algorithm: it’s sensitive to stride access pattern and can be combined into JIAJIA’s lock-based coherency protocol without much modification. The test results show that it has high prediction accuracy and achieves good overlap between communication and computation in benchmark applications. This work gives a good proof that study on software-based dynamic prefetching is reasonable, promising and worth to make further steps.