1875849 - Check if network.http.http2.enabled.deps=false hurts performance

Status: RESOLVED WORKSFORME

(Core :: Networking: HTTP, defect, P2)

Description

[Thomas Wisniewski [:twisniewski]]

As reported at webcompat.com, Firefox cannot access https://www.openrice.com/, while other browsers can. I see NS_ERROR_NET_RESET appears in the network panel for the devtools.

Comment 1

[Valentin Gosu [:valentin]]

Attachment: openrice.pcapng

I've captured the conversation to this website.
It seems the server connection is established, but then the server sends a GOAWAY frame with Error: ENHANCE_YOUR_CALM (11).
The error goes away when disabling HTTP2.

I think the error is the server's fault - we just need to prove it 🙂

Comment 2

[Valentin Gosu [:valentin]]

It seems that setting network.http.http2.enabled.deps to false makes the bug go away.
That means the server is just confused by Firefox's prioritization/depenency scheme and sending back a GOAWAY.
Some next steps here would be:

• figuring out what networking stack the server is using
• determining if this can be solved on the server side or if the impact is large enough that it would warrant a change in our HTTP/2 priority scheme.

Comment 3

[Valentin Gosu [:valentin]]

• Randell: We should see how often enhance your calm happens. I think we might have some telemetry for that already.
• Andrew: Can we do a perf experiment to see if disabling the network.http.http2.enabled.deps pref hurts performance? I expect the big servers are minding the pref. Simpler ones probably don't. Would the experiment actually be able to tell us if there's a regression? I'm thinking we should see how bug 1865040 impacts things too.
• We should figure out what server the site is using.

Comment 4

[Valentin Gosu [:valentin]]

Note that the site is currently working for me - it appears to be only using h1.1 at the moment.

It might be the proxy they're using that was causing the bug.

Server: nginx
X-Node: HKORWEB1
X-Frame-Options: sameorigin
Content-Security-Policy: frame-ancestors https://.openrice.com https://.openrice.com.cn
X-Cache-Status: Upstream from hkorrproxy1
X-Web-Version: v4
X-XSS-Protection: 1; mode=block

Comment 5

[Andrew Creskey [:acreskey]]

(In reply to Valentin Gosu [:valentin] (he/him) from comment #3)

• Andrew: Can we do a perf experiment to see if disabling the network.http.http2.enabled.deps pref hurts performance? I expect the big servers are minding the pref. Simpler ones probably don't. Would the experiment actually be able to tell us if there's a regression? I'm thinking we should see how bug 1865040 impacts things too.

I think it's a great opportunity to start looking at our prioritization systems.
Let's talk about an experiment in the wild on that flag because we've been having a hard time picking out signal from noise in telemetry experiments.
And in the meantime I'll see if network.http.http2.enabled.deps has any impact in the CI tests.

Comment 6

[Andrew Creskey [:acreskey]]

From a CI test, disabling network.http.http2.enabled.deps picked up an improvement in visual metrics on google-slides.
But this isn't a great environment to test network prioritization since there is no latency and unrealistic bandwidth.

Comment 7

[Valentin Gosu [:valentin]]

https://www.openrice.com/ seems to work now, since they only serve http/1.1
Let's use this bug to see if our dependency tree is good/bad/neutral for perf.

Comment 8

[Valentin Gosu [:valentin]]

We already set the pref off in bug 1865040