564618 - access-nbody aborts causing slowdowns, allocator thrashing

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Robert Sayre]

I noticed that we spend 6.6% of access-nbody in the kernel. The Shark Malloc Trace is pretty damning.

Comment 1

[Robert Sayre]

Attachment: Shark Malloc Trace - Chart View

Check out that sine wave at the bottom. Bad news. I'll upload the .mshark in a sec.

Comment 2

[Robert Sayre]

http://people.mozilla.com/~sayrer/2010/05/07/sunspider/mshark/access-binary-trees-malloc.mshark

Comment 3

[Brendan Eich [:brendan]]

I wonder if jemalloc does better -- not an issue once per-object scopes are gone, but worth a quick check?

/be

Comment 4

[Robert Sayre]

(In reply to comment #3)
> I wonder if jemalloc does better -- not an issue once per-object scopes are
> gone, but worth a quick check?

Scopes are the problem in access-binary-trees. 

This profile is of access-nbody, and it's mostly traffic from nanojit and js::TraceRecorder

Comment 5

[Brendan Eich [:brendan]]

Oops, I did load this in Shark, thought I saw js_GetMutableScope hot again.

Want to file dependent bugs?

/be

Comment 6

[Robert Sayre]

(In reply to comment #5)
> Oops, I did load this in Shark, thought I saw js_GetMutableScope hot again.

Oh, duh. I linked the wrong file in comment 2. that explains why you saw js_GetMutableScope.

Comment 7

[Robert Sayre]

here's the right one:

http://people.mozilla.com/~sayrer/2010/05/07/sunspider/mshark/access-nbody-malloc.mshark

Comment 8

[Ryan VanderMeulen [:RyanVM]]

Obsolete with the removal of tracejit.