Re: [quicwg/base-drafts] Spin per peer (#1982)

[mirjak <[e-mail]>]

mirjak commented on this pull request.



> @@ -244,11 +244,8 @@ on the source and destination addresses of the path,
 so that the spin bit is consistently enabled or
 disabled for repeated use of the same path.
 

I guess that depends on the heuristic you use. However, for a single flow you don't really have a recovery phase. Either you have a valid sample or not. If the base RTT didn't change, you should get a valid sample right after the re-ordering event is over. The interesting part is to observe the number of invalid samples. So for a single, long-lasting flow that you know has provided a valid spin signal previously, there is also no problem. What I'm worried about is if you monitor the invalid sample rate for an aggregate; especially if you have a lot of short flows, as you need more time to detect that a flow is spinning randomly than when using a constant value. Thus for short flows you may not be able make a decision in time and you have to filter them out. Given we have a lot of short flows on the internet, that can potentially strongly reduce your sample rate. Actually I would even prefer that disabled flows set the bit to 0 while enabled clients start with 1 because that would make it possible to have valid spin signals from he beginning. Still for any constant value, you can at least make a decision within about one RTT.

I mean this is certainly something to experiment with and the actual impact really depends on how many flows do what as well as the traffic patter of the flows. What I’m saying is that a constant value makes measurability at least a bit easier, and if there are no further concern about using a constant value, I would prefer if we recommend that.

My understanding of ekr’s initial comment was that he wanted to leave it open how to set the bit when disabled, as he didn’t see an advantage in (over-)specifying that behavior. Others were actually concerned that setting it randomly (if the randomness is not sufficiently provided) could enable fingerprinting. As such I don’t see a reason to not recommend a constant value when disabled.

-- 
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/pull/1982#discussion_r237395093