On Sat, Apr 13, 2013 at 8:40 AM, Asa Dotzler <a...@mozilla.com> wrote:

> I have a really basic question. Is PGO's performance gains something users
> are actually going to notice or are we mostly talking about synthetic
> benchmark pissing contests here?


As we saw when we accidentally disabled it the other day, it shows up all
over our performance tests.


> What's PGO's impact on start-up time or new window time or the loading of
> a Twitter or Facebook web page?
>

I don't know about Twitter or Facebook specifically, but we do have
measurements for our talos pageset.

Some examples of what we saw disabling PGO:

Ts, Paint - XP - 16.3% increase (Startup)
Ts Paint, MAX Dirty Profile - XP - 16% increase (Startup)
Ts, Paint - Win7 - 15.3% increase (Startup)
Ts Paint, MED Dirty Profile - Win7 - 15.5%  (Startup)
Ts Paint, MAX Dirty Profile - Win7 - 14% increase (Startup)

Tp5 Optimized - XP - 23.9% increase (Pageload)
Tp5 Optimized Responsiveness - XP - 42.3% increase (Pageload)
Tp5 Optimized (%CPU) - XP - 7.68% increase (Pageload)
Tp5 Optimized - Win7 - 21.9% increase (Pageload)
Tp5 Optimized Responsiveness - Win7 - 38.9% increase (Pageload)
Tp5 Optimized (%CPU) - Win7 - 5.36% increase (Pageload)

15% startup regressions and 20% pageload regressions aren't anything to
sneeze at.

I've also seen over the years that we have a really difficult time
> diagnosing and fixing some high profile crashes because of PGO and if we
> can eliminate those or make them easier to diagnose and fix and all it
> costs us is a some points on Sunspider, I'm wondering if it doesn't make
> sense to just drop PGO and focus on finding performance wins elsewhere.


Amusingly enough Sunspider is probably the least effected benchmark by PGO
because it spends all it's time in JIT code.  But if Ts and Tp5 are models
of what users see then it costs us a lot more than a few Sunspider points.
On the other hand, if they aren't good models then we essentially have no
performance testing ...

- Kyle
_______________________________________________
dev-platform mailing list
dev-platform@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-platform

Reply via email to