[
https://issues.apache.org/jira/browse/HDFS-8818?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15993654#comment-15993654
]
Kihwal Lee commented on HDFS-8818:
----------------------------------
-1. Proposing revert of this untested change.
We've run it on a small (to us) 280 node cluster. It moves a little for a
couple of iterations and hangs forever. So, contrary to the title, it becomes
extremely slow. Blindly creating a thread pool per datanode is a bad idea.
{code:java}
public void executePendingMove(final PendingMove p) {
// move the block
final DDatanode targetDn = p.target.getDDatanode();
ExecutorService moveExecutor = targetDn.getMoveExecutor();
if (moveExecutor == null) {
final int nThreads =
moverThreadAllocator.allocate(maxConcurrentMovesPerNode);
if (nThreads > 0) {
moveExecutor = targetDn.initMoveExecutor(nThreads);
}
}
if (moveExecutor == null) {
LOG.warn("No mover threads available: skip moving " + p);
return;
}
{code}
By default, this causes a pool of 50 threads to be created per target and at
some point no more can be created. If this happens, the pending move is simply
skipped. This causes eternal hang at {{waitForMoveCompletion()}}, because the
pending moves are not removed. I saw 25 thread pools with 50 threads in the
280-node cluster in this hang state.
IMO, creating more threads is not a scalable solution. Somehow allowing more
thread pool creation is definitely not a solution either.
> Allow Balancer to run faster
> ----------------------------
>
> Key: HDFS-8818
> URL: https://issues.apache.org/jira/browse/HDFS-8818
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: balancer & mover
> Reporter: Tsz Wo Nicholas Sze
> Assignee: Tsz Wo Nicholas Sze
> Fix For: 2.8.0, 2.7.4, 3.0.0-alpha1
>
> Attachments: h8818_20150723.patch, h8818_20150727.patch,
> HDFS-8818-branch-2.7.00.patch
>
>
> The original design of Balancer is intentionally to make it run slowly so
> that the balancing activities won't affect the normal cluster activities and
> the running jobs.
> There are new use case that cluster admin may choose to balance the cluster
> when the cluster load is low, or in a maintain window. So that we should
> have an option to allow Balancer to run faster.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
