[
https://issues.apache.org/jira/browse/HBASE-14708?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14980244#comment-14980244
]
Hiroshi Ikeda commented on HBASE-14708:
---------------------------------------
bq. BTW, ...
Ignore my silly opinion. I have realized NavigabaleMap is required.
I played with the benchmark code for a while.
TreeMap.floorEntry is very very light weight, and lock or synchronized is
relatively too heavy so that there is little difference when replacing
ReadWriteLock.readlock with writelock or Lock or synchronized, especially when
the map just contains about 1000 elements.
On the other hand, when putting/removing elements, ConcurrentSkipListMap is
absolutely incomparable with TreeMap with these synchronizations. Using
AtomicReference wastes too many resources when congested, and my cpus become
full-throttle in vain. Just blocking other operations is better.
> Use copy on write TreeMap for region location cache
> ---------------------------------------------------
>
> Key: HBASE-14708
> URL: https://issues.apache.org/jira/browse/HBASE-14708
> Project: HBase
> Issue Type: Improvement
> Components: Client
> Affects Versions: 1.1.2
> Reporter: Elliott Clark
> Assignee: Elliott Clark
> Priority: Critical
> Fix For: 2.0.0, 1.2.0, 1.3.0
>
> Attachments: HBASE-14708-v2.patch, HBASE-14708-v3.patch,
> HBASE-14708-v4.patch, HBASE-14708-v5.patch, HBASE-14708-v6.patch,
> HBASE-14708-v7.patch, HBASE-14708.patch, location_cache_times.pdf, result.csv
>
>
> Internally a co-worker profiled their application that was talking to HBase.
> > 60% of the time was spent in locating a region. This was while the cluster
> was stable and no regions were moving.
> To figure out if there was a faster way to cache region location I wrote up a
> benchmark here: https://github.com/elliottneilclark/benchmark-hbase-cache
> This tries to simulate a heavy load on the location cache.
> * 24 different threads.
> * 2 Deleting location data
> * 2 Adding location data
> * Using floor to get the result.
> To repeat my work just run ./run.sh and it should produce a result.csv
> Results:
> ConcurrentSkiplistMap is a good middle ground. It's got equal speed for
> reading and writing.
> However most operations will not need to remove or add a region location.
> There will be potentially several orders of magnitude more reads for cached
> locations than there will be on clearing the cache.
> So I propose a copy on write tree map.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
