623293 - AsymmGC - Thread-safe PageMap

Status: RESOLVED WONTFIX

(Tamarin Graveyard :: Garbage Collection (mmGC), defect, P3)

Description

[Simon Wilkinson]

Attachment: Initial Patch. MQ rev 181

Uniform, Tiered2 and Tiered4 are made thread-safe.
Tiered2 and Tiered4 are trivial. Uniform requires its expand operation be performed as a SafepointTask.

CacheT4 and DelayT4 are not thread-safe. Without a clever thread-safe caching scheme, it's probably not worth the overhead. (I don't have numbers or an implementation do back this up though)

Tiered4 is made the default for 64-bit builds instead of DelayT4.

Comment 1

[Simon Wilkinson]

Attachment: Latest patch. MQ rev 184

Cleaned up and fixed the Uniform build until bug 588878 lands.

Comment 2

[Simon Wilkinson]

Attachment: MQ rev 250

Comment 3

[Simon Wilkinson]

Comment on attachment 515824 [details] [diff] [review]
MQ rev 250

See http://asteam/hg/users/siwilkin/asymmGC-01 for an asymmGC patch queue
qfinished at rev 250 (against TR rev 5916).

See http://asteam.corp.adobe.com/hg/users/fklockii/tr-patch-queues for the
current queue

Comment 4

[Felix S. Klock II [:pnkfelix, :fklock]]

Performance results?  (Particularly for the 64-bit runtime variants?)

Comment 5

[Simon Wilkinson]

Attachment: Rebased. TR rev 6090. Patch queue rev 261

Comment 6

[Simon Wilkinson]

64-bit results are on the way

Comment 7

[Andre Kruetzfeldt]

Are the results available?

Comment 8

[Simon Wilkinson]

(In reply to comment #7)
> Are the results available?

As posted on the asymmGC tracking bug (bug 582770), there is a spreadsheet being maintained here:
http://asteam.corp.adobe.com/builds/downloads/siwilkin/asym_gc_perf.xlsm

This contains performance results for every patch in the asymmGC queue for Android, Windows, Linux and Mac, for both 32 and 64-bit builds where appropriate.

Comment 9

[Felix S. Klock II [:pnkfelix, :fklock]]

Comment on attachment 521938 [details] [diff] [review]
Rebased. TR rev 6090. Patch queue rev 261

(I'm still a tiny bit worried about regressions due to falling back to Tiered4 rather than DelayT4.  But this can land and I can deal with that later if need be.)

Comment 10

[William Maddox]

The following comment appears a couple of times:

    //FIXME: This critical section should be removed when the asymmGC-gcalloc
    // patch has been applied.

It's not clear exactly what patch is being referred to here, and we should make sure that the FIXME will get removed at the appropriate time.  By convention, a FIXME should include a bug number addressing the issue.  It should be sufficient to reference the bug number associated with the patch.

Comment 11

[Simon Wilkinson]

Attachment: Fixed wmaddox's points. TR rev 6090. Patch queue rev 268

Comment 12

[Dan Smith]

Retargeting to Dolores.