Improving end-to-end reliable transport using parallel transmission control protocol sessions

Achieving acceptable levels of TCP performance on high speed wide area networks with large bandwidth delay products is very difficult. Poor performance is caused by long path round trip times and non-congestive packet losses that disrupt the mechanisms used by TCP to discover available network capacity. Distributed scientific, data intensive, and high performance computing applications that require timely movement of large amounts of data over are severely hampered by the ineffectiveness of TCP. To solve performance problems, applications are increasingly relying on aggressive network protocols that can substantially improve throughput, but do so at the expense of unfairly stealing bandwidth from other applications. It is not clear that these aggressive protocols can cooperate with existing network protocols to prevent excessive network congestion or congestion collapse. This dissertation describes a new approach based on parallel TCP streams that effectively utilizes available network bandwidth, but does not unfairly appropriate bandwidth from other applications when the network is fully utilized. The approach, Combined Parallel TCP, couples one standard TCP stream with several TCP streams for which congestion avoidance has been modified to increase the number of returning acknowledgement packets required to increase the TCP congestion window. This modification exploits a characteristic of congestion avoidance in which short round trip time (RTT) TCP streams dominate long RTT TCP streams in the competition for available network bandwidth. The research findings show that Combined Parallel TCP can effectively utilize unused network bandwidth, yet prioritize fair sharing of bandwidth over effectiveness when in competition with other TCP streams. This dissertation also describes a weighted round robin packet scheduler for parallel TCP that substantially reduces the memory necessary to buffer out of sequence packets; retains the raw throughput gains of parallel TCP; and provides better goodput stability than the current unweighted round robin scheduler in use by GridFTP. Scientific, data intensive, and high performance computing applications that adopt Combined Parallel TCP and the weighted round robin scheduler described in this dissertation can effectively make use of unused network bandwidth, yet retain the critical TCP friendly and fairness characteristics necessary to fairly share the network and prevent congestion collapse.