Fine-grained load balancing with traffic-aware rerouting in datacenter networks

Abstract Modern datacenters provide a wide variety of application services, which generate mix delay-sensitive short flows and throughput-oriented long flows, transmitting in the multi-path datacenter network. Though existing load balancing designs successfully make full use available parallel paths attain high bisection network bandwidth, they reroute regardless their dissimilar performance requirements. The suffer from problems large queuing delay packet reordering, while fail to obtain throughput due low link utilization reordering. To address these inefficiency, we design fine-grained scheme, namely TR (Traffic-aware Rerouting), identifies flow types executes flexible traffic-aware rerouting balance performances both flows. Besides, avoid leverages reverse ACKs estimate switch-to-switch delay, thus excluding that potentially cause Moreover, is only deployed on switch without any modification end-hosts. experimental results large-scale NS2 simulations show reduces average tail completion time for by up 60% 80%, as well provides 3.02x gain compared state-of-the-art schemes.