I'm very interested to hear of the possible formation of a PILC group. In particular, I believe there are performance aspects peculiar to asymmetric links that need further analysis, and for which mitigations may be needed.
For instance, I've been working on HTTP performance problems over ADSL, and have uncovered previously unreported (as far as I know) interactions involving asymmetry on a link, delayed ACKs, and HTTP characteristics. These elements interact to reduce throughput significantly below expectations, to the point where it may even exceed the performance hit of ACKing every packet on an asymmetric link. This work involves analyses of packet traces of real Web downloads.
Basically, what happens is that an asymmetric link with a fast forward channel and narrow return channel (e.g., 1.5 Mbps/64Kbps) combines with HTTP characteristics (the extra upstream traffic, GET requests, rapidly increasing the modem's return channel queue length) and delayed ACK delays to increase RTT variance to the point where it often exceeds a server's RTO. The server then decides that a packet has been lost, since the RTO times out before the delayed ACK arrives. It then retransmits the "lost" packet, and slips back into slow-start mode. In the course of downloading a single 17-part page in my experiments, this happens *seven* times. Therefore, much of the transfer remains needlessly in slow-start mode. And because of all the slow-start stop-and-wait, even the narrow return channel is not fully utilized: I see an average of only 54 Kbps on the 64 Kbps channel for this download.
Reducing the asymmetry of the link alleviates this problem. I will soon be running further experiments to see whether ACKing every packet makes this situation worse or better. Hopefully, I'll be able to write this up in some coherent fashion soon and submit it to PILC, if there's interest.
I won't be able to attend the Orlando IETF meeting, but will be following the PILC developments with interest, and am interested to hear from others about experience with asymmetric link performance issues.
Thanks
= Peter
Peter G. Warren
Performance Analysis Group
Advanced Systems Laboratory
GTE Laboratories
40 Sylvan Rd., Waltham, MA 02254
(617) 466-4142
This archive was generated by hypermail 2b29 : Mon Jan 28 2002 - 09:12:19 EST