TCP Enhancements to Support High Performance Applications

Matthew Mathis

Jamshid Mahdavi

Gwendolyn L. Huntoon

Existing TCP implementations are inadequate to support high performance applications and services over national scale, high speed networks. The Pittsburgh Supercomputing Center requests funding for a project to design, develop, implement and test performance enhancements to TCP implementations. These enhancements will improve TCP based application performance over high latency, high bandwidth production networks. To accomplish this, we will focus on the Center's existing high performance computer and network environment: production supercomputers attached to the present and future wide area high speed networks.

Our ultimate goal is to eliminate TCP itself as a bottleneck, so that applications use appropriate shares of the underlying IP infrastructure with only minor latency or bandwidth penalties beyond those present in the lower layers. In other words, for all TCP connections, performance should be limited only by the application itself or congestion on an underlying network link. The only exceptions are during transients in the data rate, such as during connection startup. Furthermore it should not require network expertise to write such applications. Ultimately, all straightforward, non-privileged application codes using the standard network socket interface on a high performance computer will benefit from this work. To accomplish this goal, we propose a three year research and development project. Our primary objective is to produce an enhanced TCP that will better support high performance applications over national scale, high speed networks. We will address other important objectives and produce valuable tools for future work in this area.