L2-Cache Hierarchical Organizations for Multi-core Architectures

Abstract

Nowadays the market is moving to have multiple cores on the same chip (Chip Multiprocessors – CMP) with a multi-sliced L2 which is shared by 2 cores. CMPs with 8 cores can already be found, and future CMPs will have more than 8 cores. Typical implementations of CMPs share the L2 cache among the processors and have 2 cores sharing the same L2. We are interested in investigating the behavior of the pair: L2 sharing x L2 cache size. So, we construct models of two different organizations of CMPs: (i) tiles, with L1 and L2 private, interconnected through a router; (ii) tiles with L1 private and L2 shared among processors. The (ii) organization is evaluated with different numbers (2, 4) of cores sharing the same L2 slice and also, the L2 shared slice size is changed (1 MB, 2MB and 4 MB). With a total number of 32 cores, the proposed configurations of (ii) organization are evaluated with a full-system simulation under SPLASH-2 benchmarks. By applying both techniques, results show that the execution time is improved of about 18.9% for Ocean, 88.8% for Raytrace,and 31.8% for Volrend.