TCP Incast avoidance based on connection serialization in data center networks

Abstract

In distributed file systems, a well-known congestion collapse called TCP incast (Incast briefly) occurs because many servers almost simultaneously send data to the same client and then many packets overflow the port buffer of the link connecting to the client. Incast leads to throughput degradation in the network. In this paper, we propose three methods to avoid Incast based on the fact that the bandwidth-delay product is small in current data center networks. The first method is a method which completely serializes connection establishments. By the serialization, the number of packets in the port buffer becomes very small, which leads to Incast avoidance. The second and third methods are methods which overlap the slow start period of the next connection with the current established connection to improve throughput in the first method. Numerical results from extensive simulation runs show the effectiveness of our three proposed methods.