A GridFTP Overlay Network Service

TCP is widely deployed and is used by the current implementation of GridFTP for data transfers within grid environments. Due to TCP dynamics, throughput of data transfers can be much lower than the available bandwidth in wide area networks. Splitting a TCP connection into two or more segments has been shown to improve throughput in this setting. This paper presents a set of components to enable easy deployment of overlay networks that make use of split-TCP connections to improve GridFTP transfer performance. The components include an extension to the Globus Toolkit v4 GridFTP server that supports split TCP connections, a service to estimate bulk transfer capacity and a service to determine if and where to split a connection. Results demonstrate significant performance improvement despite using very intermittent, passive throughput observations to determine the routing of the split connections