[
https://issues.apache.org/jira/browse/HBASE-18164?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16042999#comment-16042999
]
Kahlil Oppenheimer commented on HBASE-18164:
--------------------------------------------
[~tedyu] I just made your requested changes (shortening lines, renaming, and
squashing into a single commit).
[~chia7712] The big bottleneck by far was the second bit about collecting all
the HDFS blocks of every region for every iteration of the balancer. Adding the
caching of the localities at the beginning of the balancer run is responsible
for most of the speedup.
The first part, albeit less impactful, is still important. The old locality
computation was O(# regions * # region servers), which does not scale well as
the cluster gets larger. Now it's effectively O(1), which makes a substantial
difference.
> Much faster locality cost function and candidate generator
> ----------------------------------------------------------
>
> Key: HBASE-18164
> URL: https://issues.apache.org/jira/browse/HBASE-18164
> Project: HBase
> Issue Type: Improvement
> Components: Balancer
> Reporter: Kahlil Oppenheimer
> Assignee: Kahlil Oppenheimer
> Priority: Critical
> Attachments: HBASE-18164-00.patch, HBASE-18164-01.patch,
> HBASE-18164-02.patch
>
>
> We noticed that during the stochastic load balancer was not scaling well with
> cluster size. That is to say that on our smaller clusters (~17 tables, ~12
> region servers, ~5k regions), the balancer considers ~100,000 cluster
> configurations in 60s per balancer run, but only ~5,000 per 60s on our bigger
> clusters (~82 tables, ~160 region servers, ~13k regions) .
> Because of this, our bigger clusters are not able to converge on balance as
> quickly for things like table skew, region load, etc. because the balancer
> does not have enough time to "think".
> We have re-written the locality cost function to be incremental, meaning it
> only recomputes cost based on the most recent region move proposed by the
> balancer, rather than recomputing the cost across all regions/servers every
> iteration.
> Further, we also cache the locality of every region on every server at the
> beginning of the balancer's execution for both the LocalityBasedCostFunction
> and the LocalityCandidateGenerator to reference. This way, they need not
> collect all HDFS blocks of every region at each iteration of the balancer.
> The changes have been running in all 6 of our production clusters and all 4
> QA clusters without issue. The speed improvements we noticed are massive. Our
> big clusters now consider 20x more cluster configurations.
> One design decision I made is to consider locality cost as the difference
> between the best locality that is possible given the current cluster state,
> and the currently measured locality. The old locality computation would
> measure the locality cost as the difference from the current locality and
> 100% locality, but this new computation instead takes the difference between
> the current locality for a given region and the best locality for that region
> in the cluster.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
