[
https://issues.apache.org/jira/browse/HADOOP-14937?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16197663#comment-16197663
]
Steve Loughran commented on HADOOP-14937:
-----------------------------------------
Steven: assigned to you.
Bear in mind that HTTP1.1 connections are pooled, so there will be an initial
cost of setting up an HTTP1.1 connection for that first connection (+DNS
lookup, TCP routing for long-haul); Smaller blocks may get uploaded with reuse
of the pool of http connections, eliminate setup cost
Also: the FS statistics will track #of active uploads from a stream, though
that's only tracked in stream-level stats, not in the published FS statistics
> initial part uploads seem to block unnecessarily in S3ABlockOutputStream
> ------------------------------------------------------------------------
>
> Key: HADOOP-14937
> URL: https://issues.apache.org/jira/browse/HADOOP-14937
> Project: Hadoop Common
> Issue Type: Bug
> Components: fs/s3
> Affects Versions: 3.0.0-beta1
> Reporter: Steven Rand
> Assignee: Steven Rand
> Attachments: yjp_threads.png
>
>
> From looking at a YourKit snapshot of an FsShell process running a {{hadoop
> fs -put file:///... s3a://...}}, it seems that the first part in the
> multipart upload doesn't begin to upload until n of the
> {{s3a-transfer-shared-pool}} threads are able to start uploading, where n is
> the value of {{fs.s3a.fast.upload.active.blocks}}.
> To hopefully clarify a bit, the series of events that I expected to see with
> {{fs.s3a.fast.upload.active.blocks}} set to 4 is:
> 1. An amount of data equal to {{fs.s3a.multipart.size}} is buffered into
> off-heap memory (I have {{fs.s3a.fast.upload.buffer = bytebuffer}}).
> 2. As soon as that happens, a thread begins to upload that part. Meanwhile,
> the main thread continues to buffer data into off-heap memory.
> 3. Once another part has been buffered into off-heap memory, a separate
> thread uploads that part, and so on.
> Whereas what I think the YK snapshot shows happening is:
> 1. An amount of data equal to {{fs.s3a.multipart.size}} * 4 is buffered into
> off-heap memory.
> 2. Four threads start to upload one part each at the same time.
> I've attached a picture of the "Threads" tab to show what I mean. Basically
> the times at which the first four {{s3a-transfer-shared-pool}} threads start
> to upload are roughly the same, whereas I would've expected them to be more
> staggered.
> I'm actually not sure whether this is the expected behavior or not, so feel
> free to close if this doesn't come as a surprise to anyone.
> For some context, I've been trying to get a sense for roughly which values of
> {{fs.s3a.multipart.size}} perform the best at different file sizes. One thing
> that I found confusing is that a part size of 5 MB seems to outperform a part
> size of 64 MB up until files that are upwards of about 500 MB in size. This
> seems odd, since each {{uploadPart}} call is its own HTTP request, and I
> would've expected the overhead of those to become costly at small part sizes.
> My suspicion is that with 4 concurrent part uploads and 64 MB blocks, we have
> to wait until 256 MB are buffered before we can start uploading, while with 5
> MB blocks we can start uploading as soon as we buffer 20 MB, and that's what
> gives the smaller parts the advantage for smaller files.
> I'm happy to submit a patch if this is in fact a problem, but wanted to check
> to make sure I'm not just misunderstanding something.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
