hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tomasz Nykiel (Commented) (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-2495) Increase granularity of write operations in ReplicationMonitor thus reducing contention for write lock
Date Mon, 07 Nov 2011 21:16:52 GMT

    [ https://issues.apache.org/jira/browse/HDFS-2495?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13145813#comment-13145813

Tomasz Nykiel commented on HDFS-2495:

Nicholas, there is a maximum number of blocks that can be processed at once:



    final int numlive = heartbeatManager.getLiveDatanodeCount();
    final int blocksToProcess = numlive

    workFound = this.computeReplicationWork(blocksToProcess);

So the answer to your question is no, I suppose, since we have this limit everytime the computeDatanodeWork()
is called.
> Increase granularity of write operations in ReplicationMonitor thus reducing contention
for write lock
> ------------------------------------------------------------------------------------------------------
>                 Key: HDFS-2495
>                 URL: https://issues.apache.org/jira/browse/HDFS-2495
>             Project: Hadoop HDFS
>          Issue Type: Sub-task
>          Components: name-node
>            Reporter: Tomasz Nykiel
>            Assignee: Tomasz Nykiel
>         Attachments: replicationMon.patch, replicationMon.patch-1
> For processing blocks in ReplicationMonitor (BlockManager.computeReplicationWork), we
first obtain a list of blocks to be replicated by calling chooseUnderReplicatedBlocks, and
then for each block which was found, we call computeReplicationWorkForBlock. The latter processes
a block in three stages, acquiring the writelock twice per call:
> 1. obtaining block related info (livenodes, srcnode, etc.) under lock
> 2. choosing target for replication
> 3. scheduling replication (under lock)
> We would like to change this behaviour and decrease contention for the write lock, by
batching blocks and executing 1,2,3, for sets of blocks, rather than for each one separately.
This would decrease the number of writeLock to 2, from 2*numberofblocks.
> Also, the info level logging can be pushed outside the writelock.

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


View raw message