[
https://issues.apache.org/jira/browse/SAMZA-191?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jakob Homan updated SAMZA-191:
------------------------------
Attachment: Screen Shot 2014-03-18 at 10.49.53 AM.png
VisualVM screenshot showing a fifth of the time being spent in incPoll.
Specifically, the hashmap retrieval causes lots of calls to SSP hashCode, which
was already optimized in SAMZA-106. These ConcurrentHashMaps could be tuned
(by default they're optimized for 16 concurrent writers:
http://ria101.wordpress.com/2011/12/12/concurrenthashmap-avoid-a-common-misuse/),
but I'm of the opinion that this metric is of questionable marginal value and
we should blow it away.
> Per-SSP poll count metric is overly expensive
> ---------------------------------------------
>
> Key: SAMZA-191
> URL: https://issues.apache.org/jira/browse/SAMZA-191
> Project: Samza
> Issue Type: Bug
> Components: metrics
> Affects Versions: 0.7.0
> Reporter: Jakob Homan
> Assignee: Jakob Homan
> Fix For: 0.7.0
>
> Attachments: Screen Shot 2014-03-18 at 10.49.53 AM.png
>
>
> Profiling shows that on jobs with lots of SSPs/TPs, particularly when the
> SSPs are significantly disparate in their relative volumes, due to lots of
> time incrementing the per-SSP poll count.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
