[
https://issues.apache.org/jira/browse/HDFS-918?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12831715#action_12831715
]
Jay Booth commented on HDFS-918:
--------------------------------
Absolutely -- every packet send has to fetch the BlockChannel that it's
transferring from out of a shared pool. Given that I'm using commons-pool for
this, it's probably a straight "synchronized" on every get, which would
predictably kill performance. Commons-pool is more built around the idea of a
nonsharable database connection anyways, so I'll write my own custom thingy
that works off of a ReadWriteLock and allows different threads to share the
same BlockChannel instance (since BlockChannel only exposes threadsafe channel
operations).
If that doesn't do it, well, I'll see if I can figure it out but I agree that
this is not performant enough yet. I think there's an argument for committing
this with equivalent performance because it simplifies the read pipeline on the
datanode side and consumes fewer resources (selectors/filechannels), but 1/3
the throughput, no way. On the upside, it means I won't be adding new
dependencies with this patch anymore once I rebuild that block channel pool.
(Kinda jumping the gun here on assuming the cause but it's literally the only
source of systemwide synchronization, and things only slowed down once I added
a bunch of threads, so...)
I could definitely set that random read test up, once I get the slow pool issue
sorted out. Assuming that the pool is the issue, this setup should save a
bunch of time specifically in that scenario -- no per-request buffer allocation
or file opening.
> Use single Selector and small thread pool to replace many instances of
> BlockSender for reads
> --------------------------------------------------------------------------------------------
>
> Key: HDFS-918
> URL: https://issues.apache.org/jira/browse/HDFS-918
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: data-node
> Reporter: Jay Booth
> Fix For: 0.22.0
>
> Attachments: hdfs-918-20100201.patch, hdfs-918-20100203.patch,
> hdfs-multiplex.patch
>
>
> Currently, on read requests, the DataXCeiver server allocates a new thread
> per request, which must allocate its own buffers and leads to
> higher-than-optimal CPU and memory usage by the sending threads. If we had a
> single selector and a small threadpool to multiplex request packets, we could
> theoretically achieve higher performance while taking up fewer resources and
> leaving more CPU on datanodes available for mapred, hbase or whatever. This
> can be done without changing any wire protocols.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
