Re: [quicwg/base-drafts] QUIC PTO is too conservative, causing a measurable regression in tail latency (#3526)

[ianswett <[e-mail]>]

Yes, I meant RFC 6298, and I just updated the original post to reflect that.

In terms of rttvar, the production data I'm seeing indicates rttvar is fairly close to 1/2 SRTT on average, which isn't so surprising given SRTT is an EWMA and for it to be on the order of SRTT, it would require estimates fairly close to 0 to balance out the higher estimates.

Unfortunately, rttvar is RTT **variation**, not RTT **variance**(or RTT stdev).  As such, I don't know if we can reason about what fraction of spurious PTOs would be within SRTT + 4*RTTVar, even if we could assume a normal distribution(which is also likely wrong, but I think the least problematic assumption at the moment).

I'd really rather not cap rttvar in any way, since I can't come up with any intellectual justification for that.  The 1.5 constant from TLP does seem a bit arbitrary, but it is larger than SRTT.

We added code in Chrome to try RTTstdev(EWMA'd), but that's even more experimental than the suggestion in my first post on this issue.  I think there's a good chance 2*RTTStdev is the best option(vs 4*RTTVar), but I don't have data to support that yet.

I reopened PR #2952 for packet number skipping.  I can open an issue if needed, but it's a simple non-normative PR.

-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/issues/3526#issuecomment-599833225