Active management of data caches by exploiting reuse information

As microprocessor speeds continue to outpace memory subsystems in speed, mi nimizing average data access time grows in importance. Multilateral caches afford an opportunity to reduce the average data access time by active mana gement of block allocation and replacement decisions, We evaluate and compa re the performance of traditional caches and multilateral caches with three active block allocation schemes: MAT, NTS, and PCS. We also compare the pe rformance of NTS and PCS to multilateral caches with a near-optimal, but no nimplementable policy, pseudo-opt, that employs future knowledge to achieve both active allocation and active replacement. NTS and PGS are evaluated r elative to pseudo-opt with respect to miss ratio, accuracy of predicting re ference locality, actual usage accuracy, and tour lengths of blocks in the cache. Results show the multilateral schemes do outperform traditional cach e management schemes, but fall short of pseudo-opt; increasing their predic tion accuracy and incorporating active replacement decisions would allow th em to more closely approach pseudo-opt performance.