BACKGROUND:  High speed networks with large bandwidth and long propagation delays post 2 problems for TCP performance. First, if ssthresh is set too high, compared to the path bandwidth delay product, there is the danger too many packets are generated too fast, causing multiple losses at the bottleneck router, resulting in a coarse Timeout. Or, if ssthresh is set too low, the connection exists Slow-Start prematurely, resulting in poor startup utilization. This is a particular problem when there are many short sessions, resulting in frequent Startups, e.g. HTTP.

INNOVATION:  This invention proposes a sender-side modification, called Adaptive Start (AStart). This work builds upon prior work of TCP Westwood (TCPW; UCLA Case 2003-111) in which bandwidth estimation techniques are used based on analysis of the ACK stream, bottleneck link bandwidth, buffer space at the bottleneck router, end-to-end propagation time and error rate. TCPW was a protocol improvement using bandwidth estimation to improve TCP performance of ongoing sessions in congestion or error conditions.

AStart is a protocol improvement using TCPW techniques to improve session initiation performance. AStart adaptively and repeatedly resets ssthresh when bandwidth estimation calculations indicate thee is more available capacity. By adapting to network conditions in the startup phase, the sender is able to grow the congestion window (cwnd) quickly without incurring the risk of buffer overflow and packet loss.

DEVELOPMENT TO DATE:  NS-2 simulation experiments show that AStart is able to significantly improve link utilization under various bandwidth, buffer size and round trip delay conditions.

This work was performed in the Network Research Lab at UCLA. (http://www.cs.ucla.edu/NRL/)

Reference: UCLA Case No. 2003-354

PCT Application: PCT/US04/009645