637950 - Move ArenaHeader to Chunk

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect)

Description

[Gregor Wagner [:gwagner]]

Measurements show that this accounts for many cache misses.

Comment 1

[Gregor Wagner [:gwagner]]

Attachment: patch

Comment 2

[Gregor Wagner [:gwagner]]

SS scores don't change much.
For a benchmark I create a linked list with 1000000 objects.
I measure the time[ms] to 
1) create the list
2) GC on whole list
3) delete root and GC

On trunk I get:
40, 87, 49

with this patch:
37, 77, 45

Comment 3

[Gregor Wagner [:gwagner]]

Nick do you have a way to measure cache performance with this patch?

Comment 4

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

How does it affect our V8 score?

If you're in Linux, an easy way to measure cache misses is with the perf tool. In Ubuntu, it's in the linux-tools-common package. Just run:
  perf stat ./js ...
and it will print a summary of cache stuff after the command finishes.

Comment 5

[Nicholas Nethercote [inactive]]

(In reply to comment #3)
> Nick do you have a way to measure cache performance with this patch?

Yes.  For what workload?

Comment 6

[Gregor Wagner [:gwagner]]

(In reply to comment #5)
> (In reply to comment #3)
> > Nick do you have a way to measure cache performance with this patch?
> 
> Yes.  For what workload?

SS, V8 and maybe an allocation heavy page in the browser?

Bill: I saw better scores for V8 in the browser but I still have to verify that this wasn't just noise. Haven't tried with the shell.

Comment 7

[Gregor Wagner [:gwagner]]

Attachment: patch

Comment 8

[Nicholas Nethercote [inactive]]

I tried SS and V8 in the shell, the changes all looked like noise.  I didn't try in the browser because those results tend to be much less deterministic and so very difficult to detect any meaningful difference that isn't large.