AQM with multi-queue for microburst in data center networks

Abstract

In data center networks, extremely short-term burst traffic, called microburst, causes rapid increase of queue length and accordingly severe performance degradation. Aggregated short flows generated by Partition/Aggregate is reported to be a major cause of microburst. A data center is reported to have short latency sensitive flows and long throughput sensitive flows. When congestion control reacts in early stage of queue length growth, microburst generated by short flow is expected to be resolved but throughput for long flows might be degraded. In this paper, we propose a new adaptive queue management for microburst in data center networks, MA-ECN(Microburst Aware ECN). In MA-ECN, short flow queue and long flow queue are prepared and different ECN threshold value is applied for each queue so that short queue has small threshold. Our simulation results show that MA-ECN can adequately regulate rapid queue length increase caused by microburst without throughput degradation for long flows.