Round-trip latency prediction for memory access fairness in mesh-based many-core architectures

Abstract

In mesh-based many-core architectures, processor cores and memories reside in different locations (center, corner, edge, etc.), therefore memory accesses behave differently due to their different communication distances.The latency difference leads to unfair memory access and some memory accesses with very high latencies, degrading the system performance. However, improving one memory access's latency can worsen the latency of another since memory accesses contend in the network. Therefore, the goal should focus on memory access fairness through balancing the latencies of memory accesses while ensuring a low average latency. In the paper, we address the goal by proposing to predict the round-trip latencies of memory access related packets and use the predicted round-trip latencies to prioritize the packets. The router supporting fair memory access is designed and its hardware cost is given. Experiments are carried out with a variety of network sizes and packet injection rates and prove that our approach outperforms the classic round-robin arbitration in terms of average latency and LSD. In the experiments, the maximum improvement of the average latency and the LSD are 16% and 48% respectively.