[
https://issues.apache.org/jira/browse/HBASE-5898?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13271901#comment-13271901
]
Todd Lipcon commented on HBASE-5898:
------------------------------------
Shouldn't be hard to get that CPU time back -- we can just add an array of
cacheline-padded AtomicLongs to the cache. Whenever we add something to the
cache, we do {{changeCounters[key.hashCode() %
changeCounters.length].getAndIncrement()}}. Then change the code to:
{code}
AtomicLong changeCounter = changeCounters[key.hashCode() %
changeCounters.length];
long firstTimeChangeCounter = changeCounter.get();
first time:
try to look up in cache
if found: return it
second time:
take lock:
if changeCounter.get() == firstTimeChangeCounter: it's not in cache
otherwise: look up in cache again
{code}
We'd probably want to cache-pad the AtomicLongs too to avoid false sharing.
> Consider double-checked locking for block cache lock
> ----------------------------------------------------
>
> Key: HBASE-5898
> URL: https://issues.apache.org/jira/browse/HBASE-5898
> Project: HBase
> Issue Type: Improvement
> Components: performance
> Affects Versions: 0.94.1
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Priority: Critical
> Attachments: 5898-TestBlocksRead.txt, hbase-5898.txt
>
>
> Running a workload with a high query rate against a dataset that fits in
> cache, I saw a lot of CPU being used in IdLock.getLockEntry, being called by
> HFileReaderV2.readBlock. Even though it was all cache hits, it was wasting a
> lot of CPU doing lock management here. I wrote a quick patch to switch to a
> double-checked locking and it improved throughput substantially for this
> workload.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
