[
https://issues.apache.org/jira/browse/HDFS-918?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jay Booth updated HDFS-918:
---------------------------
Attachment: hdfs-918-20100211.patch
Here's my latest patch. I replaced the dependency on commons-pool for
filechannel sharing with a ReadWriteLock-based pool where it only acquires a
global lock to register newly opened channels or while cleaning up excess
channels. I also have better integration with DataXceiverServer.
I haven't had time to benchmark and won't for a week or so, but I'm hoping for
equal speed streaming and better performance for random read.
ARCHITECTURE
* DataXCeiverServer now reads the first byte of every request (OP)
* DataXCeivers are now instantiated with an op and inputstream
* If Op was READ_BLOCK and multiplex is enabled, dispatch thread registers with
MultiplexedBlockSender and immediately dies
* MultiplexedBlockSender.register() assigns requests across a configurable pool
of SelectorThreads, round robin
* Each SelectorThread maintains selector and buffer, continually polls for
writable and writes packets, sharing resources across requests
* Maintains current wire protocol
* Packet based model adaptable to asynchronous disk I/O in the future
* Possibly adaptable for local reads by writing packets to a
WritableByteChannel wrapping a ByteBuffer
* Easy to implement connection re-use and keep-alive parameter by dumping
threads into a read Selector and re-registering or closing as necessary
> 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-918-20100211.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.
