Improving Flow Scheduling Scheme With Mix-Traffic in Multi-Tenant Data Centers

Abstract

Data centers need low-latency fabrics. Several flow scheduling schemes have been proposed to minimize the Flow Completion Time (FCT) based on Shortest Job First (SJF) heuristic. However, to mimic SJF, previous proposals sacrifice the generality (e.g., pFabric requires special hardware) or sacrifice the performance to guarantee the generality (e.g., PIAS loses some of pFabric’s performance). Especially, in multi-tenant data centers, traffic patterns from different applications are mixed together and vary over time, thereby creating even more challenges. In this paper, we investigated that the performance of information-agnostic scheme could be further improved by leveraging the unique characteristics of different traffic types. Based on this investigation, we present Traffic Prediction based Flow Scheduling (TPFS), aiming at achieving near-optimal performance and good generality in multi-tenant data centers with the mix-traffic pattern. To achieve near-optimal performance, we design a two-stage machine learning algorithm to first automatically cluster flows with the similar flow size distribution and then predict the priorities of flows based on the clustering results. Besides, we implement TPFS in virtual switches, which exerts fine-grained flow scheduling over the arbitrary network stacks of tenants. Testbed evaluation and simulations show that TPFS outperforms the previous information-agnostic flow scheduling scheme PIAS and greatly reduces the tail latency of the network.