On Wed, Oct 7, 2020 at 3:26 PM Ian Swett <ianswett= [email protected]> wrote:
> Given TCP TFO is rarely enabled, IETF QUIC without 0-RTT still typically > saves a round trip vs TCP+TLS 1.3. > Even with TLS 1.3 resumption + early data? For us the vast majority of TCP connections are resumed, meaning HTTP/2 clients can initiate its requests after 1-RTT, the same as 1-RTT QUIC. It's actually somewhat of an issue in networks with TCP PEPs since the app will think it's ready to write the first request on the TCP connection sooner than the QUIC connection, since the TCP handshake finished unnaturally early. > > The YouTube improvements are believed to be due to the lack of > retransmission ambiguity and extra ACK blocks vs TCP w/SACK. In the > presence of packet policers, both can be quite advantageous. > > Note that these are controlled experiments, so they include the metrics > from users who had QUIC enabled, but ended up using TCP due to not yet > caching Alt-Svc, UDP blockage or some other reason. > > FWIW we'll have a blog coming in the next couple weeks which also has "wow!" numbers, using similar experimental setup (though obviously with our apps instead of browsers + Alt-Svc). I would agree these are big contributors, though elimination of HOLB does measurably help connections with high degrees of multiplexing. The aforementioned PEPs are also often really terrible for application performance, despite their name suggesting otherwise. Simply bypassing these can be a big win for QUIC all by itself. > On Wed, Oct 7, 2020 at 5:42 PM Martin Duke <[email protected]> > wrote: > >> Thanks David. I'm intrigued by your performance numbers. You say there's >> no 0RTT and I've always thought of you guys running a pretty >> state-of-the-art TCP stack. Is this just HOL blocking, or is there >> something else? >> >> On Wed, Oct 7, 2020 at 9:50 AM David Schinazi <[email protected]> >> wrote: >> >>> Hi QUIC WG, >>> >>> We would like to share an important announcement from Chrome: >>> >>> https://blog.chromium.org/2020/10/chrome-is-deploying-http3-and-ietf-quic.html >>> >>> In particular, we'd like to highlight two points of interest to the WG: >>> >>> 1) Chrome now supports IETF QUIC by default (h3-29). >>> >>> 2) Since the subsequent IETF drafts 30 and 31 do not have >>> compatibility-breaking >>> changes, we currently are not planning to change the over-the-wire >>> identifier. What >>> this means is that while we'll keep tracking changes in the IETF >>> specification, we >>> will be deploying them under the h3-29/0xff00001d name. We therefore >>> recommend >>> that servers keep support for h3-29 until the final RFCs are complete if >>> they wish to >>> interoperate with Chrome. However, if the IETF were to make >>> compatibility-breaking >>> changes in a future draft, Chrome will revisit this decision. >>> >>> Full details in the link above. >>> >>> Cheers >>> David >>> >> Matt Joras
