[
https://issues.apache.org/jira/browse/HADOOP-14444?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16027771#comment-16027771
]
Lukas Waldmann commented on HADOOP-14444:
-----------------------------------------
I don't have any exact performance test. I may try to do one but generally for
filesystem with huge number of files the new implementation is way more
effective. If for nothing else than because in current implementation download
of each file requires at least 2 "connects" - if you have several thousands of
files the overhead is extraordinary. I am using the new FS mostly for distcp
and from my experience download of aprox 40000 files took about 6 hours to
even start while with new one and pooled connections it started in around 20min
add 1) it can but for another download with about 90000 files i didn't have to
fiddle with java memory parameters. In any case you can always disable it
add 2) yes but i don't think anybody sane would use ftp/sftp on "lively"
filesystem. Main point of those fs is (at least for me) use with distcp. Again
as i said you can disable caching and there is provision to figure out if a
file was changed during download
> New implementation of ftp and sftp filesystems
> ----------------------------------------------
>
> Key: HADOOP-14444
> URL: https://issues.apache.org/jira/browse/HADOOP-14444
> Project: Hadoop Common
> Issue Type: New Feature
> Components: fs
> Affects Versions: 2.8.0
> Reporter: Lukas Waldmann
> Assignee: Lukas Waldmann
> Attachments: HADOOP-14444.patch
>
>
> Current implementation of FTP and SFTP filesystems have severe limitations
> and performance issues when dealing with high number of files. Mine patch
> solve those issues and integrate both filesystems such a way that most of the
> core functionality is common for both and therefore simplifying the
> maintainability.
> The core features:
> * Support for HTTP/SOCKS proxies
> * Support for passive FTP
> * Support of connection pooling - new connection is not created for every
> single command but reused from the pool.
> For huge number of files it shows order of magnitude performance improvement
> over not pooled connections.
> * Caching of directory trees. For ftp you always need to list whole directory
> whenever you ask information about particular file.
> Again for huge number of files it shows order of magnitude performance
> improvement over not cached connections.
> * Support of keep alive (NOOP) messages to avoid connection drops
> * Support for Unix style or regexp wildcard glob - useful for listing a
> particular files across whole directory tree
> * Support for reestablishing broken ftp data transfers - can happen
> surprisingly often
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
