Transport Working Group

Phoebus is a system, developed at the University of Deladware, that seeks to improve data transport performance by splitting long network paths into distinct segments mediated by gateways that forward data between segments. By confining packet loss and latency to individual segments, Phoebus improves the end-to-end performance of applications.

The Internet2 Transport Working Group has been studying Phoebus, using the current partial Internet2 deployment, and collecting examples of applications where Phoebus significantly improves transport performance. The Working Group has found that in real world situations the use of Phoebus can improve an application's end-to-end performance experience by a factor of two or three.

GridFTP

Members of the working group have experimented with a Phoebus enabled GridFTP over 10G networks. The above graph shows four transport performance cases over a simulated connection having a 100ms WAN latency and .01% loss. Not only does Phoebus provide substantial improvement in performance for the single stream case, but a single TCP stream with Phoebus outperforms eight parallel streams without Phoebus.

REDDnet

REDDnet provides distributed storage for data intensive applications. Members of the working group have conducted tests of REDDnet transfers using Phoebus. The above graph shows the case for single thread read/write transfers between CERN and the University of Michigan. Phoebus significantly improves read performance for all transfer sizes and write transfers for sizes above 256 MB.

Interactive Visualization

One of the members of the working group has been testing Phoebus with an interactive application for dissecting anatomical data sets. While this is a low bandwidth application in comparison to bulk data transport, it also benefits from the use of Phoebus.

In this test the client, running on OS X, was located in Wisconsin, the server in California and transport occurred over WiscNet, Internet2 and CENIC with WiscNet having the greatest congestion. In response to user actions, the server transports image frames to the client. The histogram of transports versus rate shows that Phoebus increases the transport rate by a factor of three. This increase is significant and can determine whether the application is useable over a given path.