[
https://issues.apache.org/jira/browse/HDFS-946?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12832745#action_12832745
]
Hairong Kuang commented on HDFS-946:
------------------------------------
> If you are proposing that the object that is sent over-the-wire is different
> from FileStatus. If so, please consider the requirement of HDFS-878 too.
This jira tries to reduce the cost of getFileInfo and listing a directory,
where HDFS-878 adds cost to these two operations.. So I will not implement
HDFS-878 in this jira. Since we are having so many problems with getFileInfo
and list a directory, we should be very cautious about adding anything to
FileStatus in hdfs unless it is absolutely necessary.
I have conducted some experiments with my patch. I write an application that
spawns 100 threads, each of which lists a directory of size 1300 for 200 times.
I use yourKit to profile the NameNode while the application is running. Without
the patch, NameNode's CPU utilization is 20~26% and time spent on GC is 3~5%.
With the patch, NameNode's CPU utilization drops to 12~17% and the time spent
on GS is mostly 0% but occasionally becomes 1 or 2%.
> NameNode should not return full path name when lisitng a diretory or getting
> the status of a file
> -------------------------------------------------------------------------------------------------
>
> Key: HDFS-946
> URL: https://issues.apache.org/jira/browse/HDFS-946
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: Hairong Kuang
> Assignee: Hairong Kuang
> Fix For: 0.22.0
>
> Attachments: HDFSFileStatus.patch, HDFSFileStatus1.patch
>
>
> FSDirectory#getListring(String src) has the following code:
> int i = 0;
> for (INode cur : contents) {
> listing[i] = createFileStatus(srcs+cur.getLocalName(), cur);
> i++;
> }
> So listing a directory will return an array of FileStatus. Each FileStatus
> element has the full path name. This increases the return message size and
> adds non-negligible CPU time to the operation.
> FSDirectory#getFileInfo(String) does not need to return the file name either.
> Another optimization is that in the version of FileStatus that's used in the
> wire protocol, the field path does not need to be Path; It could be a String
> or a byte array ideally. This could avoid unnecessary creation of the Path
> objects at NameNode, thus help reduce the GC problem observed when a large
> number of getFileInfo or getListing operations hit NameNode.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
