1753979 - Consider increasing the number of DNS resolver threads.

Status: RESOLVED FIXED

(Core :: Networking: DNS, enhancement, P2)

Description

[Valentin Gosu [:valentin] (he/him)]

The number of parallel DNS resolutions is currently limited to 8 It has been so pretty much for since 2008.

As Brian noticed, Chrome has recently increased these limits.

This bug requires an investigation into doing the same.

Comment 1

[Bas Schouten (:bas.schouten)]

Not sure setting a perf impact is appropriate, as we can't really tell right now what the impact would be.

But umm, can't we just double the number on nightly and look at telemetry? It feels like this shouldn't be super-hard. Maybe something Marc could do? What do you think Dragana?

Comment 2

[Dragana Damjanovic [:dragana]]

Yes, we are planning to perform such an experiment.

Comment 3

[Andrew Creskey [:acreskey]]

For native dns queries, I believe the probe dns_native_queuing collects the delay.
0 ms, but 32 ms at 95th percentile.

Comment 4

[Valentin Gosu [:valentin] (he/him)]

Attachment: Bug 1753979 - Increase number of DNS resolver threads to 64 r=#necko,acreskey

The new threadpool will have 40 threads dedicated to High priority DNS
requests, and 24 for any priority DNS requests leading to a total of 64.

The recent DNS thread count experiment validates that this value is
reasonable and reduces the values of dns_native_queuing in the 95, 99 and 99.9
percentiles (all others being 0).

Moreover this has limited impact on memory usage, since this high thread
limit is only reached in times when a lot of DNS requests are made.
Normal browsing will hit this limit infrequently.

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/8a43a8a71dca
Increase number of DNS resolver threads to 64 r=necko-reviewers,kershaw

Comment 6

[Narcis Beleuzu [:NarcisB]]

https://hg.mozilla.org/mozilla-central/rev/8a43a8a71dca

Comment 7

[Andrew Creskey [:acreskey]]

Valentin shipped the configuration from treatment-a of the experiment.
From examining the nimbus experiment results (release) we can see the greatly reduced queuing time below, 95% and 99% percentiles:

control
0.95   26ms
0.99  816ms

treatment-a 
0.95    3ms
0.99  345ms

Query here.

Comment 8

[Valentin Gosu [:valentin] (he/him)]

Attachment: image.png

DNS_NATIVE_QUEUING on Nightly reflects the same improvement since we landed this.