[
https://issues.apache.org/jira/browse/DERBY-1704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12472585
]
Knut Anders Hatlen commented on DERBY-1704:
-------------------------------------------
The results shown in the attached graphs are not valid for trunk after the
commit of the DERBY-2107 patch. The single-record select load that was used is
more latch-intensive than lock-intensive, and when the latches no longer use
the lock manager, the effect of splitting the hash tables in the lock manager
is small, even with multiple CPUs. There still is a positive effect for more
lock-intensive loads though (for instance scans).
Since there are two global synchronization points in the lock manager, I'll
file subtasks so they can be discussed separately.
> Allow more concurrency in the lock manager
> ------------------------------------------
>
> Key: DERBY-1704
> URL: https://issues.apache.org/jira/browse/DERBY-1704
> Project: Derby
> Issue Type: Improvement
> Components: Performance, Services
> Affects Versions: 10.2.1.6
> Reporter: Knut Anders Hatlen
> Assigned To: Knut Anders Hatlen
> Priority: Minor
> Attachments: 1cpu.png, 2cpu.png, 8cpu.png, cleanup1.diff,
> cleanup1.stat, cleanup1.v2.diff, cleanup2.diff, cleanup3.diff, cleanup3.stat,
> split-hashtables.diff, split-hashtables.stat
>
>
> I have seen indications of severe monitor contention in SinglePool
> (the current lock manager) when multiple threads access a Derby
> database concurrently. When a thread wants to lock an object, it needs
> to obtain the monitor for both SinglePool and LockSet (both of them
> are global synchronization points). This leads to poor scalability.
> We should investigate how to allow more concurrency in the lock
> manager, and either extend SinglePool or implement a new manager.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
