On the Evaluation of Dense Chip-Multiprocessor Architectures

Chip-multiprocessors (CMPs) have been revealed as the most promising way of making efficient use of current improvements in integration scale. Nowadays, commercial CMP releases integrate at most 8 processor cores onto the chip. However, 16 or more processor cores are expected to be offered in near future dense-CMP (D-CMP) systems. In this way, these architectures impose new design restrictions, and some topics, such as the cache-coherence problem, must be reviewed. In this paper we present an exhaustive performance evaluation of two recently proposed D-CMP architectures, making special emphasis on the solution to the cache-coherence problem that each one of them introduces. The shared bus fabric architecture (SBF) features a snoop cache-coherence protocol and is based on a high-performance bus fabric interconnection network. The second architecture follows a directory-based approach and integrates a bi-dimensional mesh as the interconnection network. Our results show that the performance achieved by the SBF architecture is hard-limited by the bandwidth restrictions of the bus fabric. On the other hand, the directory-based architecture outperforms the SBF one, but presents some performance inefficiencies due to the additional indirection that the directory structure stored in the L2 cache level introduces