> TCP buffer sizing involves mainly two areas. One is good RTT > measurements to be able to estimate the bw*delay product well and the > other is information about memory (mbuf) usage in the networking > system to do the right thing if memory gets low.
Why try to measure the bw*delay? Why not use the trick from PSC's autotuning paper whereby you just try to ensure that the socket buffer size is always some multiple (2-4, I think) of the congestion window? I.e., so the congestion window dictates the performance and the socket buffer is not a factor. Of course, you have to figure out what to do to all the connections when there is not enough memory for such socket buffer sizes. But, fundementally, that seems like a much better approach to me. And, thanks for taking this all on! It sounds wonderful! allman -- Mark Allman -- ICIR -- http://www.icir.org/mallman/
pgp00000.pgp
Description: PGP signature
