End-Point Congestion Filter for Adaptive Routing with Congestion-Insensitive Performance

摘要

Interconnection networks are a critical component in most modern systems nowadays. Both off-chip networks, in HPC systems, data centers, and cloud servers, and on-chip networks, in chip multiprocessors (CMPs) and multiprocessors system-on-chip (MPSoCs), play an increasing role as their performance is vital for the performance of the whole system. One of the key components of any interconnect is the routing algorithm, which steers packets through the network. Adaptive routing algorithms have demonstrated their superior performance by maximizing network resources utilization. However, as systems increase in size (both in off-chip and on-chip), new problems emerge. One of them is congestion where traffic jams inside the network lead to low throughput and high packet latency, significantly impacting overall system performance. We propose a mechanism to eradicate this phenomena and to allow adaptive routing algorithms to achieve the expected performance even in the presence of congestion situations. End-Point Congestion Filter, EPC, detects congestion formed at the end-points of the network, and prevents the congestion from spreading through the network. Basically, EPC disables adaptivity in congested packets. Preliminary results for mid and high congestion situations show EPC is able to totally decouple congestion from routing.