[jira] [Commented] (HADOOP-12107) long running apps may have a huge number of StatisticsData instances under FileSystem

["Sangjin Lee (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-12107?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14594054#comment-14594054 ] 

Sangjin Lee commented on HADOOP-12107:
--------------------------------------

The key method in clearing this memory is in {{Statistics.visitAll()}}:

{code}
    private synchronized <T> T visitAll(StatisticsAggregator<T> visitor) {
      visitor.accept(rootData);
      if (allData != null) {
        for (Iterator<StatisticsData> iter = allData.iterator();
            iter.hasNext(); ) {
          StatisticsData data = iter.next();
          visitor.accept(data);
          if (data.owner.get() == null) {
            /*
             * If the thread that created this thread-local data no
             * longer exists, remove the StatisticsData from our list
             * and fold the values into rootData.
             */
            rootData.add(data);
            iter.remove();
          }
        }
      }
      return visitor.aggregate();
    }
{code}

As part of running the visitor, it checks to see if the underlying thread is gone, and if so, adds the data for that thread to {{rootData}} and removes the instance from the list.

This pattern almost literally cries out for using a {{PhantomReference}}. That way, we can perform this operation as soon as the garbage collector clears up the threads. I'll draw up a patch based on that idea soon.

> long running apps may have a huge number of StatisticsData instances under FileSystem
> -------------------------------------------------------------------------------------
>
>                 Key: HADOOP-12107
>                 URL: https://issues.apache.org/jira/browse/HADOOP-12107
>             Project: Hadoop Common
>          Issue Type: Bug
>          Components: fs
>    Affects Versions: 2.7.0
>            Reporter: Sangjin Lee
>            Assignee: Sangjin Lee
>            Priority: Minor
>
> We observed with some of our apps (non-mapreduce apps that use filesystems) that they end up accumulating a huge memory footprint coming from {{FileSystem$Statistics$StatisticsData}} (in the {{allData}} list of {{Statistics}}).
> Although the thread reference from {{StatisticsData}} is a weak reference, and thus can get cleared once a thread goes away, the actual {{StatisticsData}} instances in the list won't get cleared until any of these following methods is called on {{Statistics}}:
> - {{getBytesRead()}}
> - {{getBytesWritten()}}
> - {{getReadOps()}}
> - {{getLargeReadOps()}}
> - {{getWriteOps()}}
> - {{toString()}}
> It is quite possible to have an application that interacts with a filesystem but does not call any of these methods on the {{Statistics}}. If such an application runs for a long time and has a large amount of thread churn, the memory footprint will grow significantly.
> The current workaround is either to limit the thread churn or to invoke these operations occasionally to pare down the memory. However, this is still a deficiency with {{FileSystem$Statistics}} itself in that the memory is controlled only as a side effect of those operations.



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