To Improve Throughput via Multi-pathing and Parallel TCP on Each Path

Abstract

Parallel TCP, which opens multiple TCP connections over a single direct path, and Multi-Pathing, which concurrently uses multiple disjointed paths to transfer data, have both been proved to be effective methods to improve end-to-end throughput. How much throughput can we ultimately achieve between a source and a destination if we use multiple overlay paths and open multiple TCP connections on each used path? In order to find all possible overlay paths of good quality between a source and a destination, a path probing process similar to the path discovery protocol of IEEE 802.5 is started by the destination. A probing packet(a TCP connection request followed by padding data) is flooded across an overlay between the destination and the source. Intermediate overlay nodes selectively accept and forward probing packets. If a probing pack is accepted, a corresponding TCP connection is created. Trade-offs then are made between reducing the probing traffic and keeping multiple TCP connections on each path. The source strips data into small packets and adaptively assigns them to selected overlay paths according to the changing quality of each path. This proposed data transfer technology is evaluated within an overlay that consists of 15 servers on the Internet in China, across 3 different autonomous systems. Experiments show that with this technology, 54% of the measured samples yield a throughput larger than 60Mb/s, which is 60% of the bandwidth that could be possibly obtained(the access bandwidth is 100Mb/s for all servers). Comparing with direct path and Parallel TCP, only less than 1% and 25% of the measured samples reach the same level of throughput respectively.