[jboss-jira] [JBoss JIRA] (JGRP-1662) Remove flow control and fragmentation if run on TCP transport
[
https://issues.jboss.org/browse/JGRP-1662?page=com.atlassian.jira.plugin....
]
Bela Ban commented on JGRP-1662:
--------------------------------
Hmm, the perf of the old config (+TCP -UFC +MFC +FRAG2) was much better than the new one
(+TCP -UFC -MFC -FRAG2) ! In the new config, the initial few seconds were better (perf ca.
200MB/sec), but then it degraded to ca. 30MB/s !
On both the home cluster and cluster(02-5) I got ~ 88MB/sec/node with the old config and
~54MB/sec/node with the new config.
The new config had much more variation over time.
Also, with the new config, there were some sporadic *OOMEs* !
One possible explanation is that without flow control, individual nodes would send
messages as fast as TCP allowed them, leading to contention at the receivers which had to
receive data from 3 TCP pipes. With MFC, there would be less contention.
I need to investigate this, but if the results don't improve, this JIRA will be
killed.
Remove flow control and fragmentation if run on TCP transport
-------------------------------------------------------------
Key: JGRP-1662
URL: https://issues.jboss.org/browse/JGRP-1662
Project: JGroups
Issue Type: Enhancement
Reporter: Bela Ban
Assignee: Bela Ban
Fix For: 3.3.4, 3.4
When run with a TCP transport, we don't need MFC and FRAG(2). However, if we remove
FRAG2, the bundler at the transport will drop a message if it exceeds max_bundle_size.
Let's therefore bypass message bundling when the transport is TCP (TCP has its own
batching).
TODO: measure how this change affects performance on a 4 node cluster. If positive,
commit it.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira