hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Nauroth (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-7291) Persist in-memory replicas with appropriate unbuffered copy API on POSIX and Windows
Date Tue, 28 Oct 2014 03:33:34 GMT

    [ https://issues.apache.org/jira/browse/HDFS-7291?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14186292#comment-14186292

Chris Nauroth commented on HDFS-7291:

Thanks for the updated patch, Xiaoyu.  The comments in {{NativeIO}} and {{Storage}} do a great
job of explaining the whole situation.  This mostly looks good.  I have just a few more comments:

* NativeIO.java: Instead of {{org.apache.commons.io.IOUtils#closeQuietly}}, I recommend using
{{org.apache.hadoop.io.IOUtils#cleanup}}, and passing the {{LOG}} to it.  The difference is
that we'll get a log message on the rare chance that close fails.  Sometimes that extra log
message can be helpful.
* Also in NativeIO.java, the return value of {{FileChannel#transferTo}} is ignored.  Is that
intentional?  Is it not necessary to check the actual number of bytes written and loop because
we expect the destination {{FileChannel}} never uses non-blocking I/O?  If so, then it would
be worthwhile to add a comment explaining that the return value is ignored intentionally.
* NativeIO.c: The function compiles on Linux right now, because it has void return type. 
However, I think it's preferable to have a code path for Linux that calls {{THROW}} with a
"not supported" message.  We expect this code will never get called on Linux.  Right now,
if someone introduced a bug that caused Linux to call it, then it wouldn't cause an immediate
error, but it wouldn't actually do the copy, which could cause confusion.

> Persist in-memory replicas with appropriate unbuffered copy API on POSIX and Windows
> ------------------------------------------------------------------------------------
>                 Key: HDFS-7291
>                 URL: https://issues.apache.org/jira/browse/HDFS-7291
>             Project: Hadoop HDFS
>          Issue Type: Sub-task
>          Components: datanode
>    Affects Versions: 2.6.0
>            Reporter: Xiaoyu Yao
>            Assignee: Xiaoyu Yao
>         Attachments: HDFS-7291.0.patch, HDFS-7291.1.patch, HDFS-7291.2.patch
> HDFS-7090 changes to persist in-memory replicas using unbuffered IO on Linux and Windows.
On Linux distribution, it relies on the sendfile() API between two file descriptors to achieve
unbuffered IO copy. According to Linux document at http://man7.org/linux/man-pages/man2/sendfile.2.html,
this is only supported on Linux kernel 2.6.33+. 
> As pointed by Haowei in the discussion below, FileChannel#transferTo already has support
for native unbuffered IO on POSIX platform. On Windows, JDK 6/7/8 has not implemented native
unbuffered IO yet. We change to use FileChannel#transfer for POSIX and our own native wrapper
of CopyFileEx on Windows for unbuffered copy.  

This message was sent by Atlassian JIRA

View raw message