1649140 - Investigate using thread-local arenas in real-time threads

Status: NEW

(Core :: Audio/Video: cubeb, enhancement)

Description

[Paul Adenot (:padenot)]

Attachment: lockstat-threadname.py

Following my profiler work, I'm investigating lock contention on our real-time threads.

Attached is a python script (originally from https://github.com/goldshtn/linux-tracing-workshop/blob/master/lockstat-solution.py and modified) that measures the time it takes to wait for a lock, and then the time spent inside the locked region.

http://www.brendangregg.com/blog/2016-06-14/ubuntu-xenial-bcc-bpf.html has instructions to install the right packages for ubuntu, and then by running the script like so:

sudo python2 lockstat-solution.py <tab pid>

the init stacks, lock stacks, aggregated wait time, aggregated hold time, enter count are dumped on stdout each for each 5 seconds time period (roughly).

Attached is a run of https://padenot.github.io/loopa/ being profiled with the Gecko profiler, with and without calls to https://searchfox.org/mozilla-central/source/memory/build/malloc_decls.h#100 at the beginning of the AudioIPC thread and the GraphRunner thread, on a Linux build.

Comment 1

[Paul Adenot (:padenot)]

Attachment: With arenas enabled

Format is the following:

A list of init stack for each lock created.

A list of threads with their name. For each thread a list of lock acquisition during a 5s time duration, separated by stacks.

Comment 2

[Paul Adenot (:padenot)]

Attachment: Without arenas enabled (current state of things)

We see we're locking much much less with arenas enabled.

Comment 3

[Paul Adenot (:padenot)]

Attachment: Bug 1649140 - Try thread local arenas on real-time threads. DO NOT LAND

Comment 4

[Paul Adenot (:padenot)]

We can't do this right now, because there is no solution to dispose of arenas (bug 1364359), but the numbers are interesting I think and I wanted this recorded somewhere.

Comment 5

[Sean Feng [:sefeng211]]

(I know nothing about this bug)

Just FWIW, Bug 1610720 made arenas can be disposed of by calling moz_dispose_arena if the arena is free.

Comment 6

[Paul Adenot (:padenot)]

Sean, thank you very much for mentioning this, very useful. I assume we need a bit more discipline with the alloc lifetime before using this (I haven't tried yet), but it's a path forward.