[
https://issues.apache.org/jira/browse/HBASE-8517?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Elliott Clark updated HBASE-8517:
---------------------------------
Attachment: HBASE-8517-0.patch
Here's a patch that solves the issue I was seeing on a cluster.
Most of the issue was upping the move cost so that all cost functions below it
couldn't overcome the cost of a move. In other words the balancer will find
the best moves when the region load is our of wack, but it won't move just to
make locality better.
Then I also took a crack at making this thing work on larger clusters.
To do that I made it closer to simulated annealing. It picks either a random
server, or the most loaded server. This means that more often we have a chance
to work on the worst balanced server.
Then I also made the algorithm prefer to move regions off of the over loaded
servers.
Then I made the max number of moves scale with cluster size.
> Stochastic Loadbalancer isn't finding steady state on real clusters
> -------------------------------------------------------------------
>
> Key: HBASE-8517
> URL: https://issues.apache.org/jira/browse/HBASE-8517
> Project: HBase
> Issue Type: Bug
> Reporter: Elliott Clark
> Assignee: Elliott Clark
> Attachments: HBASE-8517-0.patch
>
>
> I have a cluster that runs IT tests. Last night after all tests were done I
> noticed that the balancer was thrashing regions around.
> The number of regions on each machine is not correct.
> The balancer seems to value the cost of moving a region way too little.
> {code}
> 2013-05-09 16:34:58,920 DEBUG [IPC Server handler 4 on 60000]
> org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer: Finished
> computing new load balance plan. Computation took 5367ms to try 8910
> different iterations. Found a solution that moves 37 regions; Going from a
> computed cost of 56.50254222730425 to a new cost of 11.214035466575254
> 2013-05-09 16:37:48,715 DEBUG [IPC Server handler 7 on 60000]
> org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer: Finished
> computing new load balance plan. Computation took 4735ms to try 8910
> different iterations. Found a solution that moves 38 regions; Going from a
> computed cost of 56.612624531830996 to a new cost of 11.275763861636982
> 2013-05-09 16:38:11,398 DEBUG [IPC Server handler 6 on 60000]
> org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer: Finished
> computing new load balance plan. Computation took 4502ms to try 8910
> different iterations. Found a solution that moves 39 regions; Going from a
> computed cost of 56.50048461413552 to a new cost of 11.225352339003237
> {code}
> Each of those balancer runs were triggered when there was no load on the
> cluster.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
