Adaptive thresholds are useful #3572
Conversation
|
Since RFC4653 doesn't seem like it would map to QUIC as readily as the stuff in the RACK draft about reorder-threshold tuning (it seems to be all about extending the Limited Transmit algorithm), maybe RACK is a better reference? https://tools.ietf.org/pdf/draft-ietf-tcpm-rack-08.pdf |
|
I would prefer to reference RACK, but it's not an RFC yet and I'm not sure it will be prior to QUIC shipping. What's the best way to handle that? |
|
The QUIC recovery spec already references the RACK draft in multiple places. |
|
Fair point, I replaced the old reference with the RACK reference. PTAL. |
ianswett
added
the
design
|
LG! |
ianswett
changed the title
Co-Authored-By: Martin Thomson <mt@lowentropy.net>
Co-Authored-By: Martin Thomson <mt@lowentropy.net>
|
I don't see any conversation in this PR about it, but I see a commit titled simply "non-normative" reverting the reference from RACK back to NCR. It's not a big deal, but I'm just curious about the reasoning. |
|
There's conversation on the issue(#3571 (comment)) and Jana was concerned about adding another reference to RACK as well as recommending something that we don't have a detailed algorithm(and pseudocode) for. In general, anything we have a SHOULD for has more text(and typically pseudocode) than this, so I accepted that point and adjusted the PR. For example, the pseudocode doesn't detail how to detect spurious losses, since it's expected that could be fairly complex and implementation dependent. |
ianswett
removed
the
design
|
I see, thanks for pointing me to that. But Jana's ask seems to me to be all about SHOULD versus MAY, and orthogonal to referencing NCR vs RACK; if it was a SHOULD, then you'd still have the same incompletely-specified-algorithm problem with NCR. It seems to me that now that SHOULD is removed, Jana would agree that RACK is a better reference since it maps more cleanly to QUIC. |
|
Actually, I'm not so confident that RACK maps all that cleanly either, and Jana made the point that RACK's ReoWnd adjustment algorithm hasn't been closely examined, so I'm going to bow out here- I'm happy with either NCR or RACK. |
martinthomson
added
design
Fixes #3571 by tweaking the text, but doesn't add a new SHOULD.