The transfer rates observed between ORNL and NERSC are not adequate for high-bandwith, bulk data transfers. The biggest factor in limiting transfer rates is not the available bandwidth but the "TCP slow restart" protocol. This protocol comes into play when the link is congested; once a packet is dropped subsequent packets are sent at a slowly-increasing rate. A faster restart protocol would result in improved bandwidth utilization and considerably faster transfers.

A research project to develop such a protocol - faster recovery while still being fair - is underway at ORNL. We are using a UDP client/server and adding various TCP behaviours studying the results as we go. Behaviours implemented so far are the standard TCP implementations of Slow-start, Congestion Avoidance, Fast Retransmit and Fast Recovery. We have added to this: Selective Acknowlegment (rfc2018), Data Smoothing through the use of RampDown and Forward Acknowledgment ("Forward Acknowledgment: Refining TCP Congestion Control" by Mathis & Mahdavi), and Delayed Acks.

(July 2001) The TCP-like UDP protocol has been completed and is in use in the ESnet III testing referred to above, as well as in Web100 activities. This work is being done by Tom Dunigan and Florence Fowler.

(November 2001) Using the new HPSS-HPSS version of hsi we achieve sustained bandwidths up to 12 megabytes/second.

Although we expect to give continued attention to this activity, we consider this particular project complete.

Back to Projects