hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "dhruba borthakur (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HDFS-284) dfs.data.dir syntax needs revamping: multiple percentages and weights
Date Fri, 25 Sep 2009 06:27:16 GMT

    [ https://issues.apache.org/jira/browse/HDFS-284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12759400#action_12759400
] 

dhruba borthakur commented on HDFS-284:
---------------------------------------

I agree with Paul/Allen that the current scheme of things do not work well when data directories
are not homogeneous. Can somebody list the proposal on what can be done; and once the proposal
is discussed, implement a patch to adhere to the proposal?

> dfs.data.dir syntax needs revamping: multiple percentages and weights
> ---------------------------------------------------------------------
>
>                 Key: HDFS-284
>                 URL: https://issues.apache.org/jira/browse/HDFS-284
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>         Environment: This is likely a cross-platform issue.
>            Reporter: Allen Wittenauer
>            Priority: Minor
>
> Currently, all filesystems listed in the dfs.data.dir are treated the same with respected
to the space reservation percentages.  This makes sense on homogeneous, dedicated machines,
but breaks badly on heterogeneous ones and creates a bit of a support nightmare. 
> In a grid with multiple disk sizes, the admin is either leaving space unallocated or
is required to slice up the disk.  In addition, if Hadoop isn't the only application running,
there may be unexpected collisions. In order to work around this limitation, the administrator
must specifically partition up filesystem space such that the reservation 'make sense' for
all of the configured file systems.   For example, if someone has 2 small file systems and
2 big ones on a single machine, due to various requirements (such as the OS being mirrored,
systems were built from spare parts, server consolidation, whatever).   Reserving 10% might
make sense on the small file systems  (say 7G) but 10% may leave a lot more space than desired
free on the big ones (say 50G).  
> Instead, Hadoop should support a more robust syntax for directory layout.  Ideally, an
admin should be able to specify the directory location, the amount of space reserved (in either
a percentage or a raw size syntax) for HDFS, as well as a weighting such that some file systems
may be preferred over others.  In the example above, the two larger file systems would likely
be preferred over the two smaller ones.  Additionally, the reservation on the larger file
system might be changed such that it matches the 7G on the smaller file system.
> Doing so would allow for much more complex configuration scenarios without having to
shuffle a lot of things around at the operating system level.

-- 
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