827985 - Move media/webrtc/trunk out of libxul

Status: RESOLVED WONTFIX

(Core :: WebRTC, defect, P2)

Description

[(no longer active)]

We have 2-3 more months before we become unable to link libxul on Windows.  We need to start moving things out of libxul again, and the biggest piece is media/webrtc/trunk.

It's best to do this as soon as possible, and we should treat 2 months as a "hard deadline", "hard" as in something that closes mozilla-central and causes us to consider stopping PGO on Windows...

Randell, can you please give us an estimate on when you expect this to happen?  We need to have a better plan this time compared to the previous times.

Thanks!

Comment 1

[(no longer active)]

OK, we finally hit the PGO mem limit.  Randel, when do you think you can post a patch here?  This currently blocks all development on central :(

Comment 2

[Maire Reavy [:mreavy]]

(In reply to :Ehsan Akhgari from comment #1)
> OK, we finally hit the PGO mem limit.  Randel, when do you think you can
> post a patch here?  This currently blocks all development on central :(

Who's available to help with this?  When the bug was reported 2 weeks ago, the estimate was 2 months before we got to this point (blocking development).

Comment 3

[(no longer active)]

(In reply to comment #2)
> (In reply to :Ehsan Akhgari from comment #1)
> > OK, we finally hit the PGO mem limit.  Randel, when do you think you can
> > post a patch here?  This currently blocks all development on central :(
> 
> Who's available to help with this?

What kind of help are you looking for?  I think Randel started to look into this.

> When the bug was reported 2 weeks ago, the
> estimate was 2 months before we got to this point (blocking development).

Yeah, unfortunately the problem hit us much sooner than we expected.  The patch that I landed in bug 833118 will hopefully reduce the urgency of this bug.  I'll post to dev-platform when I have confirmed that.

Comment 4

[(no longer active)]

See my latest dev-platform post.  Disabling PGO on things so far has not helped as much as I had expected.  Randell, I would really appreciate if you can continue your efforts here, as I believe we will need this patch very soon.  :(

Comment 5

[Randell Jesup [:jesup] (needinfo me)]

I'm sorry - the 600MB comment off-the-cuff in #developers was based on skimming bug 768325.  The actual number was 670MB for ALL of gkmedia, not just webrtc.  So I doubt you'll get anything useful in addition from removing webrtc/trunk from libxul.  (you'll get some, but it likely will be down in the noise compared to the numbers we need or the amount that turning off PGO for  it got us - about 200MB+).

Comment 6

[(no longer active)]

(In reply to Randell Jesup [:jesup] from comment #5)
> I'm sorry - the 600MB comment off-the-cuff in #developers was based on
> skimming bug 768325.  The actual number was 670MB for ALL of gkmedia, not
> just webrtc.  So I doubt you'll get anything useful in addition from
> removing webrtc/trunk from libxul.  (you'll get some, but it likely will be
> down in the noise compared to the numbers we need or the amount that turning
> off PGO for  it got us - about 200MB+).

Hmm, so glandium told me on IRC that according to his observations, the LTCG phase is actually the biggest memory hog.  If that's true, then splitting out webrtc code into libgkmedias will actually help.  Can we at least try and measure the difference?  If it's very small, we can leave it in...

Thanks!

Comment 7

[Mike Hommey [:glandium]]

(In reply to :Ehsan Akhgari from comment #6)
> (In reply to Randell Jesup [:jesup] from comment #5)
> > I'm sorry - the 600MB comment off-the-cuff in #developers was based on
> > skimming bug 768325.  The actual number was 670MB for ALL of gkmedia, not
> > just webrtc.  So I doubt you'll get anything useful in addition from
> > removing webrtc/trunk from libxul.  (you'll get some, but it likely will be
> > down in the noise compared to the numbers we need or the amount that turning
> > off PGO for  it got us - about 200MB+).
> 
> Hmm, so glandium told me on IRC that according to his observations, the LTCG
> phase is actually the biggest memory hog.  If that's true, then splitting
> out webrtc code into libgkmedias will actually help.  Can we at least try
> and measure the difference?  If it's very small, we can leave it in...

If you compare the current setup, where webrtc has NO_PROFILE_GUIDED_OPTIMIZE=1 with webrtc out of libxul, i doubt it will make much difference. As I said on irc, NO_PROFILE_GUIDED_OPTIMIZE=1 disables both PGO and LTCG.

Comment 8

[(no longer active)]

(In reply to Mike Hommey [:glandium] from comment #7)
>  As I said on irc, NO_PROFILE_GUIDED_OPTIMIZE=1 disables
> both PGO and LTCG.

Oh, sorry, this is what I was missing.