[
https://issues.apache.org/jira/browse/HDFS-9700?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15128971#comment-15128971
]
Colin Patrick McCabe commented on HDFS-9700:
--------------------------------------------
Good find, [~ghelmling].
bq. [~iwasakims] wrote: Adding configuration key such as
HdfsClientConfigKeys.DFS_CLIENT_SOCKET_TCP_NODELAY might be conservative option
to retain existing behaviour and change the default value later. (You can see
HDFS-8829 and HDFS-9259 as example for the fix.)
Yeah, it makes sense to have a separate configuration key controlling whether
{{TCP_NODELAY}} is set on {{DataTransferProtocol}}.
I think we should change the default to be that TCP_NODELAY is "on" for both
{{DataTransferProtocol}} and Hadoop RPC. We already try to avoid sending small
messages over DataTransferProtocol, so Nagle's algorithm doesn't add a lot (and
may significantly degrade the performance of things like hflush and hsync).
> DFSClient and DFSOutputStream do not respect TCP_NODELAY config in two spots
> ----------------------------------------------------------------------------
>
> Key: HDFS-9700
> URL: https://issues.apache.org/jira/browse/HDFS-9700
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Affects Versions: 2.7.1, 2.6.3
> Reporter: Gary Helmling
> Assignee: Gary Helmling
> Attachments: HDFS-9700-v1.patch, HDFS-9700_branch-2.7.patch
>
>
> In {{DFSClient.connectToDN()}} and
> {{DFSOutputStream.createSocketForPipeline()}}, we never call
> {{setTcpNoDelay()}} on the constructed socket before sending. In both cases,
> we should respect the value of ipc.client.tcpnodelay in the configuration.
> While this applies whether security is enabled or not, it seems to have a
> bigger impact on latency when security is enabled.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
