Re: [quicwg/base-drafts] QPACK improvement: wrap absolute index values (#1644)

[Mike Bishop <[e-mail]>]

I think doubling-or-more is attempting to guard against a situation where X was valid pre-wrap but about to expire, then enough entries are added that X is valid again post-wrap, and there's confusion about whether a new frame refers to the old or new one, correct?  If so, the existing prohibition on evicting entries which still have references should ensure that there's no confusion, and the fact that real entry sizes will produce more of a gap in real usage further insulates us from the problem.  If it's intended to prevent a different issue, please clarify.

Your math about how much it costs is correct, but somewhat misconstrues the idea -- the window is being expanded to the rest of the byte *instead of* being doubled or more, not expanded after doubling.  For smaller table sizes, that's more conservative.

So for example, modulo any off-by-one errors in counting:

| MaxEntries | Expansion | Doubled |
|---|---|---|
| 128 | [0..254) = 1 byte | (254 * 1 + 2 * 2) / 256 = 1.008 bytes |
| 256 | [0..382) = (254 * 1 + 128 * 2) / 382 = 1.335 bytes | [0..512) = (254 * 1 + 128 * 2 + 130 * 3) = 1.76 bytes |
| 2000 | [0..16639) = (254 * 1 + 128 * 2 + 16257 * 3) = 2.96 bytes | [0..4000) = (254 * 1 + 128 * 2 + 3618 * 3) = 2.81 bytes

However, you're right that it becomes more tilted to the largest possible encoding once you pass two bytes, so that's not ideal for very large table sizes.

-- 
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/1644#issuecomment-411869426