JMX monitoring is extremely memory inefficient

[janssk77]

Hi,

I'm using jboss EAP 6.1.1. When reading any JMX attribute(s), I have noticed a noticeable impact on the performance of our system. After attaching a profiler, I see that getting a single attribute of a JMX bean (either using remoting or using jconsole) is generating thousands of objects, that have to cleaned up by the GC. For instance, 6000 instances of the ModelNode class are created.

A clone is made of the complete management model before executing any JMX operation. This happens in the OperationContextImpl.readResourceFromRoot class.

I wonder whether there is a workaround for this. IMHO, JMX monitoring should be as light as possible..

[dlofthouse]

Could you please confirm which attribute you are reading when you see this?  Is this the attribute from an MBean or a JMX representation of the management model?

[ctomc]

Also can you try with EAP 6.3 as there ware lots of changes in that area around 6.2+ times.

[janssk77]

I'm reading an attribute from the jboss.as domain. it does not really matter which one as long as it comes from the jboss.as domain.

eg: attribute serverState of jboss.as:management-root=server

I have not tried with another server, but looking at the latest code in github, the clone still happens.

The first thing the MBean server does is finding a delegate. In that process, it already creates a clone of the whole management model. Once it finds the delegate, it gets the actual attribute, creating another clone of the mgnt model..

[ctomc]

And again I would ask you if you can test with newer code as this part has changed considerably.

Try with either EAP 6.3 or with WildFly 8.1 or even better WildFly 9 alpha1 we released last week.

[janssk77]

Tried with wildfly 9 alpha1

The cloning of the management model still happens. Since the management model of an empty jboss is not that big, the overhead is small.

However, once more things are configured (queues/datasources), I'm afraid the cloning will become a bottleneck again...

[ctomc]

Any chance you can file bug report in WildFly jira for this https://issues.jboss.org/browse/WFLY

[janssk77]

Done:

[WFLY-3889] JMX monitoring is extremely memory inefficient - JBoss Issue Tracker

[ctomc]

Proposed solution https://github.com/wildfly/wildfly-core/pull/199