Improving TCP Throughput for AIX Hosts Communicating Over High Latency Network Connections

This document describes a method for improving the Transmission Control Protocol (TCP) throughput between AIX hosts communicating over a high latency network connection (e.g., a wide area network (WAN)). MITRE observed that the TCP throughput between AIX hosts communicating across a high bandwidth WAN was low while that of Linux hosts was significantly higher. While researching this issue, MITRE learned that AIX uses the NewReno TCP congestion algorithm, and that Red Hat Enterprise Linux (RHEL) uses the Cubic TCP congestion algorithm. These algorithms are responsible for the observed difference in TCP throughput across the WAN.

To improve the throughput for AIX hosts, a method was developed using RHEL hosts and a utility called “socat” that allows the Linux hosts to act as TCP proxies for the AIX hosts’ WAN communications. Since Linux uses the Cubic algorithm for TCP congestion control, the overall TCP throughput is improved over NewReno. MITRE observed between seven and 10 times more throughput between AIX hosts using this method.