Talos Regression :( Tp5 (RSS) increase 19.4% on MacOSX 10.5.8 Firefox
Nobody
---------------------------------------------------------------
Previous: avg 176938400.000 stddev 3183489.785 of 30 runs up to revision 763ff2d737e7
New : avg 211186200.000 stddev 3133763.903 of 5 runs since revision 1ad1fd67e97a
Change : +34247800.000 (19.4% / z=10.758)
Graph : http://mzl.la/oX09g4
Changeset range: http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=763ff2d737e7&tochange=1ad1fd67e97a
Changesets:
* http://hg.mozilla.org/mozilla-central/rev/2b2f584dc5fd
: Paul Biggar <pbi...@mozilla.com>, Jason Evans <jas...@canonware.com> and Dale Kim <dale...@illinois.edu> - Bug 414946 (part 1): Fix jemalloc on Mac, but leave disabled (r=pavlov)
Overwrite the OSX default zone allocator, taking into account the malloc_zone_t
version (supported versions are 3, 6, and 8) for Leopard, Snow Leopard and
Lion. jemalloc can be dynamically disabled for unknown malloc_zone_t versions,
for OSX 10.8 and beyond.
The changeset does not enable jemalloc, to allow for easy disabling if there's
a problem. It will be enabled in the next changeset.
This should be a 15-20% improvement in memory usage.
: http://bugzilla.mozilla.org/show_bug.cgi?id=414946
* http://hg.mozilla.org/mozilla-central/rev/1ad1fd67e97a
: Paul Biggar <pbi...@mozilla.com> - Bug 414946 (part 2): Enable jemalloc on Mac (r=pavlov)
: http://bugzilla.mozilla.org/show_bug.cgi?id=414946
Nobody
--------------------------------------------------------
Previous: avg 419.267 stddev 5.635 of 30 runs up to revision 763ff2d737e7
New : avg 394.792 stddev 3.986 of 5 runs since revision 1ad1fd67e97a
Change : -24.475 (5.84% / z=4.343)
Graph : http://mzl.la/nw2HP8
Nobody
-----------------------------------------------------------------
Previous: avg 107.717 stddev 1.112 of 30 runs up to revision 763ff2d737e7
New : avg 104.000 stddev 0.500 of 5 runs since revision 1ad1fd67e97a
Change : -3.717 (3.45% / z=3.343)
Graph : http://mzl.la/nDkj2j
Nobody
------------------------------------------------------------------
Previous: avg 3361.157 stddev 19.821 of 30 runs up to revision 763ff2d737e7
New : avg 3484.760 stddev 11.845 of 5 runs since revision 1ad1fd67e97a
Change : +123.603 (3.68% / z=6.236)
Graph : http://mzl.la/nyfGr3
Nobody
-------------------------------------------------------------------
Previous: avg 239.782 stddev 1.724 of 30 runs up to revision 763ff2d737e7
New : avg 233.861 stddev 0.685 of 5 runs since revision 1ad1fd67e97a
Change : -5.920 (2.47% / z=3.435)
Graph : http://mzl.la/rmIAt8
Nobody
------------------------------------------------------------------
Previous: avg 3514.155 stddev 31.355 of 30 runs up to revision 763ff2d737e7
New : avg 3668.596 stddev 19.765 of 5 runs since revision 1ad1fd67e97a
Change : +154.441 (4.39% / z=4.926)
Graph : http://mzl.la/ruky8K
Kyle Huey
nob...@cruncher.build.sjc1.mozilla.com> wrote:
> Regression :( Tp5 (RSS) increase 19.4% on MacOSX 10.5.8 Firefox
> ---------------------------------------------------------------
> Previous: avg 176938400.000 stddev 3183489.785 of 30 runs up to revision
> 763ff2d737e7
> New : avg 211186200.000 stddev 3133763.903 of 5 runs since revision
> 1ad1fd67e97a
> Change : +34247800.000 (19.4% / z=10.758)
> Graph : http://mzl.la/oX09g4
>
> Changeset range:
> http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=763ff2d737e7&tochange=1ad1fd67e97a
>
> Changesets:
> * http://hg.mozilla.org/mozilla-central/rev/2b2f584dc5fd
> : Paul Biggar <pbi...@mozilla.com>, Jason Evans <jas...@canonware.com>
> and Dale Kim <dale...@illinois.edu> - Bug 414946 (part 1): Fix jemalloc
> on Mac, but leave disabled (r=pavlov)
>
> Overwrite the OSX default zone allocator, taking into account the
> malloc_zone_t
> version (supported versions are 3, 6, and 8) for Leopard, Snow Leopard and
> Lion. jemalloc can be dynamically disabled for unknown malloc_zone_t
> versions,
> for OSX 10.8 and beyond.
>
> The changeset does not enable jemalloc, to allow for easy disabling if
> there's
> a problem. It will be enabled in the next changeset.
>
> This should be a 15-20% improvement in memory usage.
> : http://bugzilla.mozilla.org/show_bug.cgi?id=414946
>
> * http://hg.mozilla.org/mozilla-central/rev/1ad1fd67e97a
> : Paul Biggar <pbi...@mozilla.com> - Bug 414946 (part 2): Enable
> jemalloc on Mac (r=pavlov)
> : http://bugzilla.mozilla.org/show_bug.cgi?id=414946
>
> Bugs:
> * http://bugzilla.mozilla.org/show_bug.cgi?id=414946
> _______________________________________________
> dev-tree-management mailing list
> dev-tree-...@lists.mozilla.org
> https://lists.mozilla.org/listinfo/dev-tree-management
>
We expected a 15-20% improvement and got a 20% regression?
- Kyle
Paul Biggar
about 20 non-spikes. Am I right in thinking that those 6 were the only
ones which tested this jemalloc changesets?
I only tested this on 10.6 I confess. I'll look at the 10.6 graphs
shortly to make sure that it did have the desired effect. It's pretty
easy to turn this off on 10.5 only.
Another possibility is that I'm aware of the RSS increase, but the
measure we actually care about isnt RSS. I remember something like
this from my notes, but I'll have to look again.
Paul
--
Paul Biggar
Compiler Geek
pbi...@mozilla.com
@paulbiggar
Paul Biggar
> I only tested this on 10.6 I confess. I'll look at the 10.6 graphs
> shortly to make sure that it did have the desired effect. It's pretty
> easy to turn this off on 10.5 only.
Just to clarify, I tested the 32-bit version on 10.6, not on 10.5. But
I dont think that's the problem, see below.
> Another possibility is that I'm aware of the RSS increase, but the
> measure we actually care about isnt RSS. I remember something like
> this from my notes, but I'll have to look again.
My notes on performance are at
https://bugzilla.mozilla.org/show_bug.cgi?id=414946#c101. In
particular, I noted at the time that jemalloc doesn't actively return
memory to the OS until it is asked to, which means that RSS doesn't go
down until something else asks for that memory. Does Tp5 have
something trying to reclaim that memory? If not, then this increase is
not a real problem, but a problem with how/what we measure (unless my
assumptions are wrong).
The raw numbers looked like this:
Number are in the form: Full load -> closed tabs ->
after a while
Mac, 32bit, Private mem: 464 -> 130 -> 120
jemalloc, 32bit, Private mem: 426 -> 288 -> 91
Mac, 32bit, Real mem: 579 -> 420 -> 402
jemalloc, 32bit, Real mem: 581 -> 446 -> 207
"Real mem" is RSIZE (which I'm pretty sure is RSS), "Private mem" is
RPRVT (RSIZE - shared memory).
The process to measure this was described in the comment, but in summary:
1. open a ton of tabs, then measure "Full load"
2. close those tabs and wait a short while, then measure "closed tabs"
3. Launch a background process which uses tons of memory, then measure
"after a while".
So what's apparent from this is that the memory usage only drops with
jemalloc after another process asks for it. Is the way I've measured
correct, and if so, should we ignore the Tp5 regression? (CCing njn
for a hopefully definitive answer).
Paul
PS Due to bug 670175 I think we'll be disabling jemalloc soon anyway.
But I'd love to figure this out definitively for when we (hopefully)
reenable it.
Nicholas Nethercote
<n.neth...@gmail.com> wrote:
>
> It sounds plausible; AIUI Tp5 (RSS) is measured at 20 second
> intervals and then all the measurements are averaged, so if jemalloc
> is holding onto things longer it could increase the measurements.
It would be instructive to compared Linux and Windows runs with and
without jemalloc enabled, to see if the same thing happens.
Nick
Nicholas Nethercote
>
> So what's apparent from this is that the memory usage only drops with
> jemalloc after another process asks for it. Is the way I've measured
> correct, and if so, should we ignore the Tp5 regression? (CCing njn
> for a hopefully definitive answer).
It sounds plausible; AIUI Tp5 (RSS) is measured at 20 second
intervals and then all the measurements are averaged, so if jemalloc
is holding onto things longer it could increase the measurements.
Nick
Paul Biggar
Backing out shortly.
On Fri, Jul 8, 2011 at 20:07, Nicholas Nethercote
<n.neth...@gmail.com> wrote:
> On Sat, Jul 9, 2011 at 1:06 PM, Nicholas Nethercote
> <n.neth...@gmail.com> wrote:
>>
>> It sounds plausible; AIUI Tp5 (RSS) is measured at 20 second
>> intervals and then all the measurements are averaged, so if jemalloc
>> is holding onto things longer it could increase the measurements.
>
> It would be instructive to compared Linux and Windows runs with and
> without jemalloc enabled, to see if the same thing happens.
>
> Nick
>
--