Proximity coherence for chip-multiprocessors

Abstract

Many-core architectures provide an efficient way of harnessing the growing numbers of transistors available in modern fabrication processes; however, the parallel programs run on these platforms are increasingly limited by the energy and latency costs of communication. Existing designs provide a functional communication layer but do not necessarily implement the most efficient solution for chip-multiprocessors, placing limits on the performance of these complex systems. In an era of increasingly power limited silicon design, efficiency is now a primary concern that motivates designers to look again at the challenge of cache coherence. The first step in the design process is to analyse the communication behaviour of parallel benchmark suites such as Parsec and SPLASH-2. This thesis presents work detailing the sharing patterns observed when running the full benchmarks on a simulated 32-core x86 machine. The results reveal considerable locality of shared data accesses between threads with consecutive operating system assigned thread IDs. This pattern, although of little consequence in a multi-node system, corresponds to strong physical locality of shared data between adjacent cores on a chip-multiprocessor platform. Traditional cache coherence protocols, although often used in chip-multiprocessor designs, have been developed in the context of older multi-node systems. By redesign- ing coherence protocols to exploit new patterns such as the physical locality of shared data, improving the efficiency of communication, specifically in chip-multiprocessors, is possible. This thesis explores such a design – Proximity Coherence – a novel scheme in which L1 load misses are optimistically forwarded to nearby caches via new dedicated links rather than always being indirected via a directory structure.