[
https://issues.apache.org/jira/browse/HIVE-10128?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gopal V updated HIVE-10128:
---------------------------
Attachment: hashmap-after.png
Fix looks good, the hashmap sync points have gone away from the inner loops.
!hashmap-after.png!
The orange sections are the IO elevator lagging behind the map-join operator.
Found other init-time lock sections, will file more JIRAs.
> LLAP: BytesBytesMultiHashMap does not allow concurrent read-only access
> -----------------------------------------------------------------------
>
> Key: HIVE-10128
> URL: https://issues.apache.org/jira/browse/HIVE-10128
> Project: Hive
> Issue Type: Sub-task
> Reporter: Gopal V
> Assignee: Sergey Shelukhin
> Attachments: HIVE-10128.patch, hashmap-after.png,
> hashmap-sync-source.png, hashmap-sync.png
>
>
> The multi-threaded performance takes a serious hit when LLAP shares
> hashtables between the probe threads running in parallel.
> !hashmap-sync.png!
> This is an explicit synchronized block inside ReusableRowContainer which
> triggers this particular pattern.
> !hashmap-sync-source.png!
> Looking deeper into the code, the synchronization seems to be caused due to
> the fact that WriteBuffers.setReadPoint modifies the otherwise read-only
> hashtable.
> To generate this sort of result, run LLAP at a WARN log-level, to avoid all
> the log synchronization that otherwise affects the thread sync.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
