[
https://issues.apache.org/jira/browse/HDFS-8287?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14968160#comment-14968160
]
Jing Zhao commented on HDFS-8287:
---------------------------------
Thanks for updating the patch, [~lewuathe]. Also thanks for the review,
[~rakeshr].
For the current patch, looks like we're not currently using the two buffers
included in {{DoubleCellBuffer}}. Specifically, {{writeParityCells}} does the
following in a serial manner:
# wait for the encoding work (against the ready buffer) to finish
# write encoded data in the ready buffer to all the data streamers
# flip buffers
# start encoding in the new ready buffer
Since #3 and #4 happen only after #2 finishes, the above steps actually only
need/utilize one buffer. Maybe what we can do is:
# move the above #2 to the end
# the checksum calculation can be done along with encoding
Besides, it will be nice to run some simple performance tests to see what
improvement the changes can achieve.
> DFSStripedOutputStream.writeChunk should not wait for writing parity
> ---------------------------------------------------------------------
>
> Key: HDFS-8287
> URL: https://issues.apache.org/jira/browse/HDFS-8287
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: hdfs-client
> Reporter: Tsz Wo Nicholas Sze
> Assignee: Kai Sasaki
> Attachments: HDFS-8287-HDFS-7285.00.patch,
> HDFS-8287-HDFS-7285.01.patch, HDFS-8287-HDFS-7285.02.patch,
> HDFS-8287-HDFS-7285.03.patch, HDFS-8287-HDFS-7285.04.patch,
> HDFS-8287-HDFS-7285.05.patch, HDFS-8287-HDFS-7285.06.patch,
> HDFS-8287-HDFS-7285.07.patch, HDFS-8287-HDFS-7285.08.patch,
> HDFS-8287-HDFS-7285.09.patch, HDFS-8287-HDFS-7285.10.patch,
> HDFS-8287-HDFS-7285.11.patch, HDFS-8287-HDFS-7285.WIP.patch,
> HDFS-8287-performance-report.pdf, HDFS-8287.12.patch, HDFS-8287.13.patch,
> HDFS-8287.14.patch, HDFS-8287.15.patch, h8287_20150911.patch, jstack-dump.txt
>
>
> When a stripping cell is full, writeChunk computes and generates parity
> packets. It sequentially calls waitAndQueuePacket so that user client cannot
> continue to write data until it finishes.
> We should allow user client to continue writing instead but not blocking it
> when writing parity.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
