Backoff Algorithms to Avoid TCP Incast in Data Center Networks

When a standard TCP implementation using the minimum retransmission timeout (RTOmin) of 200 ms is used in distributed file systems in data centers, a well-known throughput collapse called TCP Incast occurs in barrier synchronized applications, because 200 ms is too large as an RTOmin in data center networks. A TCP implementation (TCPFG) using a much smaller RTOmin attained by a fine-grained kernel timer is proposed in order to avoid TCP Incast. However, throughput still decreases in TCPFG when the number of servers in distributed file systems is large. In order to improve the performance of TCPFG, we have proposed a method (NOBTCP) which use a new TCP option which forces servers belonging to a barrier synchronized application to retransmit packets. Although the performance of the method is much better, it is practically hard to introduce the new TCP option. In this paper, in order to improve the performance of TCPFG, we propose using different backoff algorithms from the binary exponential backoff algorithm in TCPFG. Numerical examples show that TCPFG with each of the proposed backoff algorithms shows better performance than NOBTCP.