harmony-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pavel Ozhdikhin (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HARMONY-5396) [drlvm][em] excess ValueProfiler locking causes degradation on JITted code with profiling
Date Tue, 12 Feb 2008 12:57:08 GMT

    [ https://issues.apache.org/jira/browse/HARMONY-5396?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12568094#action_12568094

Pavel Ozhdikhin commented on HARMONY-5396:

Alexey, for the long-running workloads the precision of the new ValueProfile will count and
we need to compare new strategies here as well. Neither of 4 committed techniques guarantees
precise value profile. Thus removing original locking approach is questionable to me. Only
check on real benchmarks may answer.

Regarding the value profiling overhead  - it was noticeable on DaCapo which is not a long-running
benchmark. To heavily exercise value profiler we need to turn on value profile guided devirt
for all methods, including virtual ones. VP-guided devirt for virtual calls was disabled in
server mode just because it had a noticeable profiling overhead.

Egor, in general I agree that branches are not as much affect performance as locks. The question
is do we need to keep so many profiling options? My understanding on the dev list was that
we experiment with the startegies  and choose 1-2 best to commit. And I'm concerned about
removal of precise profile locking (until measurements show that it is needless).

> [drlvm][em] excess ValueProfiler locking causes degradation on JITted code with profiling
> -----------------------------------------------------------------------------------------
>                 Key: HARMONY-5396
>                 URL: https://issues.apache.org/jira/browse/HARMONY-5396
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>            Reporter: Aleksey Shipilev
>            Assignee: Egor Pasko
>         Attachments: 0001-ValueProfileCollector-refactoring-to-separate-FirstN-and-Divided-methods.txt,
0002-value-profiling-options-locked-flagged_all-flagged_insert-unsafe.txt, 0002-ValueProfileCollector-implementing-locking-strategies.txt,
0003-value-profiling-options-locked-flagged_all-flagged_insert-unsafe.txt, HARMONY-5396-syncTie.patch
> During investigation of serious slowdown in some microbenchmark, excess locking in ValueProfiler
was identified. 
> Please see Harmony-dev thread for the details: http://markmail.org/message/tatsfbyb7eb3q3rr

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message