You are viewing a plain text version of this content. The canonical link for it is here.
Posted to common-dev@hadoop.apache.org by "Hairong Kuang (JIRA)" <ji...@apache.org> on 2009/02/03 02:21:59 UTC

[jira] Updated: (HADOOP-5124) A few optimizations to FsNamesystem#RecentInvalidateSets

     [ https://issues.apache.org/jira/browse/HADOOP-5124?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Hairong Kuang updated HADOOP-5124:
----------------------------------

    Attachment: optimizeInvalidate2.patch

This patch incorporates Konstantin's first comment.

As for (3), I am not clear how to evaluate the optimization. The goal of the new strategy is not to improve performance. Instead it aims to give fairness to all nodes when computing invalidation work. The current strategy always favors the ones in the beginning of the map since recentInvalidateSets is a TreeMap so it is sorted. Another flaw is that after the first node is scheduled, the node is reinserted into the map if it still has remaining blocks. Since it becomes the first node again, next call of invalidateWorkForOneNode will work on the same node again. The current strategy would work fine if recentInvalidateSets is a FIFO queue.

> A few optimizations to FsNamesystem#RecentInvalidateSets
> --------------------------------------------------------
>
>                 Key: HADOOP-5124
>                 URL: https://issues.apache.org/jira/browse/HADOOP-5124
>             Project: Hadoop Core
>          Issue Type: Improvement
>          Components: dfs
>            Reporter: Hairong Kuang
>            Assignee: Hairong Kuang
>             Fix For: 0.21.0
>
>         Attachments: optimizeInvalidate.patch, optimizeInvalidate1.patch, optimizeInvalidate2.patch
>
>
> This jira proposes a few optimization to FsNamesystem#RecentInvalidateSets:
> 1. when removing all replicas of a block, it does not traverse all nodes in the map. Instead it traverse only the nodes that the block is located.
> 2. When dispatching blocks to datanodes in ReplicationMonitor. It randomly chooses a predefined number of datanodes and dispatches blocks to those datanodes. This strategy provides fairness to all datanodes. The current strategy always starts from the first datanode.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.