[
https://issues.apache.org/jira/browse/HADOOP-713?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12542616
]
dhruba borthakur commented on HADOOP-713:
-----------------------------------------
We experimented on a system that has 6 million files. A "du /" with recursion
implemented on the namenode took less than 2 seconds. There was no measure-able
change in CPU usage or GC on the nameode when this command was run.
A "lsr /" on the same cluster took 13 minutes. The CPU on the namenode
increased from its normal of 4% to 10%. We saw garbage collection occuring but
the Eden-Heap-Space on the namenode remained within its normal limits.
Given the above, I propose that we accept this patch into 0.15 and trunk so
that the immediate performance bottleneck on the namenode is fixed. I will
create a new JIRA that describes that du should be implemented at the namenode
using one RPC call.
> dfs list operation is too expensive
> -----------------------------------
>
> Key: HADOOP-713
> URL: https://issues.apache.org/jira/browse/HADOOP-713
> Project: Hadoop
> Issue Type: Improvement
> Components: dfs
> Affects Versions: 0.8.0
> Reporter: Hairong Kuang
> Assignee: dhruba borthakur
> Priority: Blocker
> Fix For: 0.15.1
>
> Attachments: optimizeComputeContentLen.patch
>
>
> A list request to dfs returns an array of DFSFileInfo. A DFSFileInfo of a
> directory contains a field called contentsLen, indicating its size which
> gets computed at the namenode side by resursively going through its subdirs.
> At the same time, the whole dfs directory tree is locked.
> The list operation is used a lot by DFSClient for listing a directory,
> getting a file's size and # of replicas, and getting the size of dfs. Only
> the last operation needs the field contentsLen to be computed.
> To reduce its cost, we can add a flag to the list request. ContentsLen is
> computed If the flag is set. By default, the flag is false.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.