Searching an Optimal History Size for History-Based Page Prefetching on Software DSM Systems

This work presents a study regarding the search for an optimal value of the history size for the prediction/prefetching technique history-based prefetching, which collects the recent history of accesses to individual shared memory pages and uses that information to predict the next access to a page. On correct predictions, this technique allows to hide the latency generated by page faults on the remote node when the access is effectively done. Some parameters as the size of the page history structure that is stored and transmitted among nodes can be fine-tuned to improve the prediction efficency. Our experiments show that small values of history size provide a better performance in the tested applications, while bigger values tend to generate more latency when the page history is transmitted, without improving the prediction efficiency.