hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Steve Loughran (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-5556) A few improvements to DataNodeCluster
Date Tue, 24 Mar 2009 19:59:52 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-5556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12688835#action_12688835
] 

Steve Loughran commented on HADOOP-5556:
----------------------------------------

#Does CreateEditsLog have to create a default Configuration() instance, or could createEditsLog()
be changed to take a Configuration as a parameter? 
# It might be useful to think -at some point in time- to move these MiniDFSCluster and  DataNodeCluster
classes into a redistributable JAR, as they are handy to anyone trying to set up a test cluster
in a single JVM -though we'd have to get rid of all emergency System.exit() calls first.

> A few improvements to DataNodeCluster
> -------------------------------------
>
>                 Key: HADOOP-5556
>                 URL: https://issues.apache.org/jira/browse/HADOOP-5556
>             Project: Hadoop Core
>          Issue Type: Bug
>          Components: test
>            Reporter: Hairong Kuang
>             Fix For: 0.21.0
>
>         Attachments: DataNodeCluster.patch
>
>
> DataNodeCluster is a great tool to simulate a large scale DFS cluster using a small set
of machines. A few suggestions to improve this tool:
> # DataNodeCluster uses MiniDFSCluster#startDataNode to start multiple instances of DataNode
on one machine. MiniDFSCluster sets DataNode's address to be 127.0.0.1. We should allow to
set its address to 0.0.0.0 so DataNodes in different machines could communicate.
> # Currently the size of the blocks injected to DataNode and created in CreatedEditsLog
is hardcoded as 10. It would be more convenient if this could be configurable. Also we need
to make sure that both use the same block size.
> # If the replication factor of blocks is larger than 1, currently a DataNode in DataNodeCluster
will be injected blocks multiple times and therefore it sends block reports to NameNode multiple
times. Initial block reports contain only a portion of its blocks and therefore may cause
unnecessary block replications. It would be cleaner if only one block report with all its
blocks is sent. 

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


Mime
View raw message