545334 - 4k size buffer is very small and makes page loading/files downloading noticeable slower

Status: RESOLVED FIXED

(Core :: Networking, defect)

Description

[Oleg Romashin (:romaxa)]

We did some measurements on Maemo devices with different NECKO buffers size, SMALL 2k, normal 4k and increased(16k).
And found that with 16k buffers we have less SYS-CPU load (kernel/drivers).

It is easy to check even with wget, that downloading with small buffer size (2-4k) make CPU load higher that with 16k

Comment 1

[Oleg Romashin (:romaxa)]

Attachment: netwerk buffer size fix

Comment 2

[Oleg Romashin (:romaxa)]

Attachment: Patch for downloading

Comment 3

[Doug Turner (:dougt)]

we probably can just completely remove the necko small buffers stuff.  No one uses it (or should given this new data).

What is the ideal values here?  Can we run different values through tp4?

Comment 4

[Doug Turner (:dougt)]

Comment on attachment 426171 [details] [diff] [review]
netwerk buffer size fix

need data.

Comment 5

[Doug Turner (:dougt)]

i'll clean up the code, and get some data.

Comment 6

[Doug Turner (:dougt)]

filed: https://bugzilla.mozilla.org/show_bug.cgi?id=545869

Comment 8

[Doug Turner (:dougt)]

Alfred suggested that we keep this bug open for the FF preferences that would change the buffer size and segment count

Comment 9

[Doug Turner (:dougt)]

Attachment: Tp4 Data

I tested some builds using different buffer sizes and segment counts.  It looks likes 24 buffers of 32k should be the default used in necko.  What is more interesting is that CPU usage is much less than the other runs.

Comment 10

[Doug Turner (:dougt)]

Attachment: patch v.1

Comment 11

[Johnny Stenback (:jst)]

Seems like a worthy thing to do for 1.9.3, given the performance improvements.

Comment 12

[Doug Turner (:dougt)]

http://hg.mozilla.org/mozilla-central/rev/a8c70cdbf50b

will close if this needs to be backed out for tests.

Comment 13

[Doug Turner (:dougt)]

sad... we don't run tp4_%cpu on linux against mc; only on try.  worse, perftastic isn't showing anything conclusive on WINNT5.1 for tp4(cpu%).  Either way, it didn't hurt anything making this change.  Marking fixed.

I filed a bug 567304 to get tp4_%cpu run on linux.