hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Esteban Gutierrez (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-8564) BlockPoolSlice.checkDirs() will trigger excessive IO while traversing all sub-directories under finalizedDir
Date Mon, 22 Feb 2016 21:50:18 GMT

    [ https://issues.apache.org/jira/browse/HDFS-8564?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15157764#comment-15157764
] 

Esteban Gutierrez commented on HDFS-8564:
-----------------------------------------

I had patched Hadoop 2.6 with HDFS-8845 in order to avoid running into this problem and I
still noticed some IO spikes. I remember I even tuned vm.vfs_cache_pressure to try to cache
as much as possible all inode and dentries but just the current directory traversal is too
expensive when we haven't cached that dir or if the cache has been evicted. For sure I can
confirm that HDFS-8845 has been helpful and the frequency of the spikes has reduced but scanning
up to 64k directories can be really IO intensive.



> BlockPoolSlice.checkDirs() will trigger excessive IO while traversing all sub-directories
under finalizedDir
> ------------------------------------------------------------------------------------------------------------
>
>                 Key: HDFS-8564
>                 URL: https://issues.apache.org/jira/browse/HDFS-8564
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: datanode
>    Affects Versions: 3.0.0
>            Reporter: Esteban Gutierrez
>            Assignee: Esteban Gutierrez
>            Priority: Critical
>
> DataNodes continuously call checkDiskErrorAsync() for multiple operations in the DN in
order to verify if a volume hasn't experienced any failure. When DN.startCheckDiskErrorThread()
is invoked we need to traverse all configured data volumes on a DN to see which volumes need
to be removed (see FsVolumeList.checkDir(s)) however that means that for each directory on
BlockPoolSlice  we need to call DiskChecker.checkDirs() which will recursively will look into
the rbw, tmp and finalized directories:
> {code}
> void checkDirs() throws DiskErrorException {
>     DiskChecker.checkDirs(finalizedDir);
>     DiskChecker.checkDir(tmpDir);
>     DiskChecker.checkDir(rbwDir);
>   }
> {code}
> Unfortunately after HDFS-6482, the subdirectory structure is created with the following
algorithm:
> {code}
> public static File idToBlockDir(File root, long blockId) {
>     int d1 = (int)((blockId >> 16) & 0xff);
>     int d2 = (int)((blockId >> 8) & 0xff);
>     String path = DataStorage.BLOCK_SUBDIR_PREFIX + d1 + SEP +
>         DataStorage.BLOCK_SUBDIR_PREFIX + d2;
>     return new File(root, path);
>   }
> {code}
> Which leaves each data volume with  64K directories (256 directories x 256 subdirectories)
A side effect of this is that if the dentries haven't been cached by the OS, then the DN needs
to recursively scan up to 64k directories x the number of configured data volumes (x number
of files) impacting IO for other operations while DiskChecker.checkDirs(finalizedDir) is running.
> There are few possibilities to address this problem:
> 1. Do not scan at all finalizedDir
> 2. Limit to one level the number of sub directories to scan recursively. (256)
> 3. Remove a subdirectory immediately it doesn't have any block under it. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message