[
https://issues.apache.org/jira/browse/HBASE-15047?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15074590#comment-15074590
]
Duo Zhang commented on HBASE-15047:
-----------------------------------
OK, I should add a limitation 'in production'. In micro benchmark, yes, it can
be much faster. But in production environment, as [~stack] said above, thread
scheduling is not controlled by us so we can hardly tell how long a spin will
continue thus the spin may waste lots of CPU time.
And see this paper
"Ad Hoc Synchronization Considered Harmful"
https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Xiong.pdf
The key point here is "Ad Hoc Synchronization" is not easy to understand so
people will easily break your initial assumption and cause a very bad
performance impact.
Thanks.
> Try spin lock for MVCC completion
> ---------------------------------
>
> Key: HBASE-15047
> URL: https://issues.apache.org/jira/browse/HBASE-15047
> Project: HBase
> Issue Type: Bug
> Reporter: Elliott Clark
> Assignee: Elliott Clark
> Attachments: HBASE-15047.patch
>
>
> Waits/Notify is very very expensive since it can cost a thread scheduling.
> There should only ever be a few threads ( < Num Cores ) running. So it should
> be possible to spin and use compare and swap to update the read point.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
