[
https://issues.apache.org/jira/browse/STORM-886?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14741636#comment-14741636
]
ASF GitHub Bot commented on STORM-886:
--------------------------------------
Github user zhuoliu commented on the pull request:
https://github.com/apache/storm/pull/700#issuecomment-139676554
Initial tests on 5 node openstack cluster (1 zk, 1 nimbus, 3 supervisors).
1. Test a normal topology (no congestion happen)
WordCountTopology3 (3 times more workers and executors than original
WordCount).
Without topology running, the zookeeper node's received packets is 5
pack/second;
without BP, the zk load is 20.1 pack/second;
with BP, the zk load is 20.9 pack/second.
This makes sense: if a topology is running fluently (no congestion
happens), the ZK will almost never be accessed by Backpressure procedures.
"almost" is because we have added an additional topo-backpressure recurring
checking with credential thread (just for dealing with the very rare case that
ZK callback fails to proceed), which reads from ZK every 30 seconds.
So 12 workers cause 12/30 = 0.4 pack / second overheads to ZK.
This shows that Backpressure will have minimal additional overheads for any
non-congested topologies.
> Automatic Back Pressure
> -----------------------
>
> Key: STORM-886
> URL: https://issues.apache.org/jira/browse/STORM-886
> Project: Apache Storm
> Issue Type: Improvement
> Reporter: Robert Joseph Evans
> Assignee: Zhuo Liu
> Attachments: an simple example for backpressure.png, backpressure.png
>
>
> This new feature is aimed for automatic flow control through the topology DAG
> since different components may have unmatched tuple processing speed.
> Currently, the tuples may get dropped if the downstream components can not
> process as quickly, thereby causing a waste of network bandwidth and
> processing capability. In addition, it is difficult to tune the
> max.spout.pending parameter for best backpressure performance. Therefore, an
> automatic back pressure scheme is highly desirable.
> Heron proposed a form of back pressure that does not rely on acking or max
> spout pending. Instead spouts throttle not only when max.spout.pending is
> hit, but also if any bolt has gone over a high water mark in their input
> queue, and has not yet gone below a low water mark again. There is a lot of
> room for potential improvement here around control theory and having spouts
> only respond to downstream bolts backing up, but a simple bang-bang
> controller like this is a great start.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
