DeTail: Reducing the Flow Completion Time Tail in Datacenter Networks

摘要

Web sites are increasingly backed by complex processing to deliver rich content to users via web pages. Despite the increased complexity, the pages must still be delivered quickly and consistently to meet user's expectations for interactivity. To achieve this, data centers typically employ application-level mechanisms to squeeze in as much complex processing as possible while still meeting page delivery deadlines. However, network variability can result in variable packet latency and a long flow completion time tail. This ultimately leads to either reduced page quality to meet deadlines or increased deadline misses. In this paper, we evaluate the benefits of a new network congestion management approach for reducing the flow completion time tail. We argue that in-network traffic management, multipath data transfers, and traffic differentiation are essential for reducing the tail. We validate our approach through DeTail, an in-network multipath-aware congestion management mechanism that reduces the flow completion tail, increasing the likelihood that complex web sites will be able to meet interactive deadlines. We show that DeTail effectively reduces the 99th percentile flow completion tail for a wide range of steady and bursty workloads by up to 80%.