431221 - Disable glib slice allocator

Status: RESOLVED FIXED

(Core :: Memory Allocator, defect)

Description

[Jason Evans]

Attachment: tp (G_SLICE=always-malloc)

glib provides a slice API[1], which is implemented as a zone allocator layered
on top of malloc.  Since jemalloc is itself a zone allocator, the extra
abstraction level forces segregation of equal-sized objects that could
otherwise be packed together.

I ran the talos benchmark and plotted the memory usage for tp (see attachment).  Mapped memory (tp_Private Bytes) is markedly lower if the slice allocator is disabled.  Resident set size (tp_RSS) also appears to be slightly lower, on the order of 1-2%.  A longer run may show more significant fragmentation reduction.

[1] http://library.gnome.org/devel/glib/stable/glib-Memory-Slices.html

Comment 1

[Jason Evans]

Attachment: Disable glib slice allocator

Comment 2

[Doug Turner (:dougt)]

Comment on attachment 318233 [details] [diff] [review]
Disable glib slice allocator

this looks like the right thing, but bsmedberg is the guy to review it.

does /dev/null always exist on every unix platform we care about?

Do all application on linux using jemalloc use run-mozilla.sh (xulrunner, ff, tb?)  I am wondering if we can do this inside of the app programmatically?  (probably too late?)

Comment 3

[Daniel Brooks [:db48x]]

Instead of using >/dev/null, simply close the fd with >-

Comment 4

[Mats Palmgren (inactive)]

Could we use 'sh' builtin commands to avoid starting a new process?
For example:
: ${G_SLICE:=always-malloc}

Comment 5

[Justin Dolske [:Dolske]]

Just match existing usage in run-mozilla.sh with

if [ -z "$G_SLICE" ]
then
  G_SLICE=always-malloc
  export G_SLICE
fi

Comment 6

[Benjamin Smedberg]

Comment on attachment 318233 [details] [diff] [review]
Disable glib slice allocator

yeah, use test -z "$G_SLICE"

But, is there really no way to do this programmatically from main()?

Comment 7

[Jason Evans]

Attachment: Disable glib's slice allocator

Re: [ -z "$G_SLICE" ], I specifically avoided that construct because it overrides an empty, but existing, G_SLICE environment variable.  My intention was to be minimally disruptive.

That said, the attached patch unconditionally disables the slice allocator programmatically.

Comment 8

[timeless]

i know that some people like trying to make /dev/zero readonly and it seems some people also like to make /dev/null readonly. not sure if you care. I do know that we've had talkback crash reports (i.e. user complains talkback is broken) for one of those two cases (I can't remember which).

Comment 9

[Jason Evans]

changeset:   15459:621becf19fe6
user:        Jason Evans <[e-mail]>
date:        Fri Jun 20 10:34:42 2008 -0700
summary:     Bug 431221: Disable glib slice allocator, r=benjamin

Comment 10

[Zack Weinberg (:zwol)]

(In reply to comment #8)
> i know that some people like trying to make /dev/zero readonly and it seems
> some people also like to make /dev/null readonly. 

offtopic, but doing either is completely deranged and not worth supporting; more so for the latter; i'm amazed the system even boots with a read-only null device.