[
https://issues.apache.org/jira/browse/HDFS-16292?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17448595#comment-17448595
]
tomscut commented on HDFS-16292:
--------------------------------
Finally, by fetching *strace* on the client, we can see that the timeout
parameter is indeed set successfully.
The problem we encountered was that the pressure at the DataNode was high and
the lock competition was fierce (we had not merged this patch
[HDFS-13359|https://issues.apache.org/jira/browse/HDFS-13359] yet), resulting
in the extremely slow {*}epoll_wait{*}, which reached the level of seconds, so
the DataNode response was very slow.
!image-2021-11-24-21-04-26-568.png|width=700,height=302!
In addition, spark sets a very small buffer size (512 bytes) to read ORC files,
resulting in slow file reading. Makes people think it's infinite waiting.
!image-2021-11-24-21-06-49-172.png|width=701,height=127!
> The DFS Input Stream is waiting to be read
> ------------------------------------------
>
> Key: HDFS-16292
> URL: https://issues.apache.org/jira/browse/HDFS-16292
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: datanode
> Affects Versions: 2.5.2
> Reporter: Hualong Zhang
> Priority: Minor
> Attachments: HDFS-16292.path, image-2021-11-01-18-36-54-329.png,
> image-2021-11-02-08-54-27-273.png, image-2021-11-24-21-04-26-568.png,
> image-2021-11-24-21-06-26-064.png, image-2021-11-24-21-06-49-172.png
>
>
> The input stream has been waiting.The problem seems to be that
> BlockReaderPeer#peer does not set ReadTimeout and WriteTimeout.We can solve
> this problem by setting the timeout in BlockReaderFactory#nextTcpPeer
> Jstack as follows
> !image-2021-11-01-18-36-54-329.png!
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
