High-Performance Networks: Optimizing Quality Of Service Using TCP Auto Tuning

Abstract

TCP has become the dominant protocol for all network data transport because it presents a simple uniform data delivery service. With the growth of high performance networking a single host may have simultaneous connections that vary in bandwidth. Consequently, connections on a single host can and will scale in bandwidth by several orders. TCP, now, requires queues proportional in size to a path’s bandwidth. Traditional statically fixed mechanisms for allocating and limiting TCP queue space fall short on today’s Internet, and often limit throughput only a small fraction of the available bandwidth specially for long- fat network (large bandwidths and high round-trip times) Subsequent work on TCP has enabled the use of larger flow-control windows, yet the use of these options is still relatively rare, because manual tuning has been required. Other work has developed means for avoiding this manual tuning step, but those solutions lack generality and exhibit unfair characteristics. We have identified requirements for an automatically tuning TCP to achieve maximum throughput across all connections simultaneously within the resource limits of the sender. The aim of the thesis work is to upgrade the TCP implementation of Linux kernel 2.6.26.5 to remove above barriers while using memory efficiently. This technique results in greatly improved performance, a decrease in packet loss under bottleneck conditions, and greater control of buffer utilization by the end hosts.