A Probabilistic Approach to Address TCP Incast in Data Center Networks

Abstract

Data centers typically host tens of thousands of servers that communicate with each other using high speed network interconnects. While these servers help in servicing millions of clients, their overall performance largely depends on the efficiency of the center's communication fabric. Cost and compatibility reasons however, persuade many data centers to consider Ethernet for their baseline communication fabric. Until recently, Ethernet speeds inside data centers averaged around 100Mbps but the evolution of IEEE 802.3 standards has led to the development of 1 Gbps and 10 Gbps Ethernet. This sudden jump in Ethernet speeds requires proportional scaling of TCP/IP processing for network intensive applications to really benefit from the increased bandwidth. While IP is expected to scale well in this context, TCP is known to have problems supporting very high data rates at very low latencies. One such problem, termed the `Incast', results in gross under-utilization of link capacity in certain many-to-one TCP communication patterns. This paper presents a practical solution to TCP's incast problem. Our proposed technique relies on a probabilistic approach that augments TCP's standard congestion recovery mechanism. Simulation results demonstrate that this technique is effective in avoiding TCP throughput collapse in data center networks.