[xcp] Notes from XCP meeting at ISI
M.Handley at cs.ucl.ac.uk
Fri Feb 20 15:34:25 PST 2004
>At 5:48 PM -0500 2/20/04, Tim Shepard wrote:
>>I see this as a problem, but am unsure what to do about it. My hope
>>that someone clever can figure out how to let XCP co-exist with
>>traffic in non-XCP queues by using the dropped packets as a strong
>>hint that there's a non-XCP queue on the path which is currently the
>>bottleneck, and somehow behaving appropriately in that case.
>>I hope this inspires someone to invent a better scheme.
>>I'm guessing that a better scheme will make the TTL scheme unnecessary.
>What he said.
>we talked about this a while ago at ISI, without a lot of comfort.
>But I like Tim's idea, conceptually. I think you can state the
>approach crisply as "construct a model of the system assuming it's
>all XCP, and then watch for behaviors that are deviations from that
>model". (There are probably others than just loss). The more I think
>about it, the more I think you could go a long way with this.
The issue is clearly wireless. We'd really like to work better than
TCP on lossy wireless links with link-layer retransmit and other
curious artifacts. In principle we can, because XCP's explicit
feedback will drive our rate back up again rapidly, especially if the
wireless router knows what's going on and compensates by increasing
the amount of positive feedback it sends. But then this directly
conflicts with implicit non-XCP router discovery. Sigh.
More information about the xcp