Improving flow completion time for short flows in datacenter networks

Abstract

Today's cloud datacenters host wide variety of applications which generate diverse mix of internal datacenter traffic. In a cloud datacenter environment 90% of the traffic flows, though they constitute only 10% of the data carried around, are short flows with sizes up to a maximum of 1MB. The rest 10% constitute long flows with sizes in the range of 1MB to 1GB. Throughput matters for long flows whereas short flows are latency sensitive. Datacenter Transmission Control Protocol (DCTCP) is a transport layer protocol that is widely deployed at datacenters nowadays. DCTCP aims to reduce the latency for short flows by keeping the queue occupancy at the datacenter switches under control while ensuring throughput requirements are met for long flows. But, DCTCP congestion control algorithm treats short flows and long flows equally. We demonstrate that treating them differently, by reducing the congestion window for short flows at a lower rate compared to long flows at the onset of congestion, we could improve the flow completion time for short flows by up to 25%, thereby reducing their latency up to 25%. We have implemented a modified version of DCTCP for cloud datacenters, based on the DCTCP patch available for Linux, which achieves better flow completion time for short flows while ensuring that throughput of long flows are not affected.