Agent-Based Memory Access for Many-Core CMPs

The trend of increasing on-chip core counts and integrating memory controllers (MCs) makes core-to-memory communication a major obstacle in scaling memory access performance for many-core CMPs. Unmanaged on-chip traffic for long-distance memory accesses, combined with information asymmetry between cores and remote MCs may lead to serious inefficiency in processing massive parallel memory accesses. In this paper we propose a novel agent-based memory access model for CMPs. We employ multiple agents inside the network to assist memory accesses to remote MCs, whose role lies in conducting memory requests from nearby cores, merging some repetitive memory requests and optimizing the scheduling under the backpressure of target MCs. We further describe a simple but effective case for agent-based memory access called Quad Agent, which deploys static agent modules in each quadrant of the network to serve requests towards MCs in other quadrants. The details of memory access merging, scheduling schemes and the architectural supports are discussed. Simulation results show that Quad Agent can reduce memory access latency by 13.3% on average and achieve 20.5% IPC speedup compared with the baseline. The performance promotion is due to increased memory access merge rate, row buffer hit rate and also prevention of traffic congestion and bank starvation.