[
https://issues.apache.org/jira/browse/HADOOP-11446?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14259682#comment-14259682
]
Thomas Demoor commented on HADOOP-11446:
----------------------------------------
Hi Ted,
Looks good to me. Some minor remarks:
- The parameters should be defined (and documented) in Constants.java. Your
default of fs.s3a.threads.core=256 means up to 256 parallel (part)Uploads. That
should fill up your network pipe :p. However, the number of concurrent http
connections opened by the underlying AmazonS3Client (fs.s3a.max.connections) is
set to a much lower value by default (too low?). Could you elaborate on the
default values? I think we should tweak these a bit to give a good "out of the
box" experience and/or document some tuning guidelines for different network
conditions (use cases).
- Also use the shiny new single TransferManager for purging at the end of
initialize() in S3AFileSystem, replacing the following code path
{code}
if (purgeExistingMultipart) {
TransferManager transferManager = new TransferManager(s3);
{code}
- I like that you went for a low-level implementation for the Executor instead
of using Executors.newFixedThreadPool. The ability to block submitting threads
by setting fs.s3a.max.total.tasks is nice tool for limiting memory
consumption. Out of curiosiity: can you envision use cases where setting
different values for core.threads and max.threads would be important?
> S3AOutputStream should use shared thread pool to avoid OutOfMemoryError
> -----------------------------------------------------------------------
>
> Key: HADOOP-11446
> URL: https://issues.apache.org/jira/browse/HADOOP-11446
> Project: Hadoop Common
> Issue Type: Bug
> Reporter: Ted Yu
> Assignee: Ted Yu
> Attachments: hadoop-11446-001.patch
>
>
> Here is part of the output including the OOME when hbase snapshot is exported
> to s3a (nofile ulimit was increased to 102400):
> {code}
> 2014-12-19 13:15:03,895 INFO [main] s3a.S3AFileSystem: OutputStream for key
> 'FastQueryPOC/2014-12-11/EVENT1-IDX-snapshot/.hbase-snapshot/.tmp/EVENT1_IDX_snapshot_2012_12_11/
> 650a5678810fbdaa91809668d11ccf09/.regioninfo' closed. Now beginning upload
> 2014-12-19 13:15:03,895 INFO [main] s3a.S3AFileSystem: Minimum upload part
> size: 16777216 threshold2147483647
> Exception in thread "main" java.lang.OutOfMemoryError: unable to create new
> native thread
> at java.lang.Thread.start0(Native Method)
> at java.lang.Thread.start(Thread.java:713)
> at
> java.util.concurrent.ThreadPoolExecutor.addWorker(ThreadPoolExecutor.java:949)
> at
> java.util.concurrent.ThreadPoolExecutor.execute(ThreadPoolExecutor.java:1360)
> at
> java.util.concurrent.AbstractExecutorService.submit(AbstractExecutorService.java:132)
> at
> com.amazonaws.services.s3.transfer.internal.UploadMonitor.<init>(UploadMonitor.java:129)
> at
> com.amazonaws.services.s3.transfer.TransferManager.upload(TransferManager.java:449)
> at
> com.amazonaws.services.s3.transfer.TransferManager.upload(TransferManager.java:382)
> at
> org.apache.hadoop.fs.s3a.S3AOutputStream.close(S3AOutputStream.java:127)
> at
> org.apache.hadoop.fs.FSDataOutputStream$PositionCache.close(FSDataOutputStream.java:72)
> at
> org.apache.hadoop.fs.FSDataOutputStream.close(FSDataOutputStream.java:106)
> at org.apache.hadoop.io.IOUtils.copyBytes(IOUtils.java:54)
> at org.apache.hadoop.io.IOUtils.copyBytes(IOUtils.java:112)
> at org.apache.hadoop.fs.FileUtil.copy(FileUtil.java:366)
> at org.apache.hadoop.fs.FileUtil.copy(FileUtil.java:356)
> at org.apache.hadoop.fs.FileUtil.copy(FileUtil.java:356)
> at org.apache.hadoop.fs.FileUtil.copy(FileUtil.java:338)
> at
> org.apache.hadoop.hbase.snapshot.ExportSnapshot.run(ExportSnapshot.java:791)
> at org.apache.hadoop.util.ToolRunner.run(ToolRunner.java:70)
> at
> org.apache.hadoop.hbase.snapshot.ExportSnapshot.innerMain(ExportSnapshot.java:882)
> at
> org.apache.hadoop.hbase.snapshot.ExportSnapshot.main(ExportSnapshot.java:886)
> {code}
> In S3AOutputStream#close():
> {code}
> TransferManager transfers = new TransferManager(client);
> {code}
> This results in each TransferManager creating its own thread pool, leading to
> the OOME.
> One solution is to pass shared thread pool to TransferManager.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
