Throughput optimization of TCP incast congestion control in large-scale datacenter networks

Abstract

The many-to-one traffic pattern in datacenter networks leads to Transmission Control Protocol (TCP) incast congestion and puts unprecedented pressure to cloud service providers. The abnormal TCP behaviors in incast increase system response time and unavoidably reduce the applicability of cloud-based system deployments. This paper proposes Receiver-oriented Congestion Control (RCC) to address heavy incast in large-scale datacenter networks. RCC is motivated by oscillatory queue size of switch when handling heavy incast and substantial potential of receiver in congestion control when using TCP. RCC makes effective use of centralized scheduler and Explicit Congestion Notification (ECN) at receiver. The RCC prototype is realized in network simulator 3 (ns3) which implements TCP exactly. This paper details the RCC design and evaluates its performance in diverse and heavy workloads. The evaluation results indicate that RCC has an average decreases of 47.5% in the mean queue size and 51.2% in the 99th-percentile latency in the heavy incast over TCP.