[mod_cluster-issues] [JBoss JIRA] (MODCLUSTER-322) Using AverageSystemLoadMetric can improperly cause a Load Factor of 0

[Aaron Ogburn (JIRA)]

Aaron Ogburn created MODCLUSTER-322:
---------------------------------------

             Summary: Using AverageSystemLoadMetric can improperly cause a Load Factor of 0
                 Key: MODCLUSTER-322
                 URL: https://issues.jboss.org/browse/MODCLUSTER-322
             Project: mod_cluster
          Issue Type: Bug
    Affects Versions: 1.2.1.Final, MOD_CLUSTER_1_0_10_GA_CP02
         Environment: *JBoss Enterprise Application Platform 5
*Apache httpd
*mod_cluster 1.0.10.GA_CP02 or 1.2.1
            Reporter: Aaron Ogburn
            Assignee: Jean-Frederic Clere
             Fix For: MOD_CLUSTER_1_0_10_GA_CP03


It looks like AverageSystemLoadMetric is not properly implemented.  When using it, mod_cluster may always report a load factor of 0, thus making the JBoss node unreachable from Apache.  We've tested with a simple web app that checks the underlying MXBean SystemLoadAverage:

Double.class.cast(server.getAttribute(ObjectName.getInstance(ManagementFactory.OPERATING_SYSTEM_MXBEAN_NAME), "SystemLoadAverage")).doubleValue()


This info is grabbed pretty much the same way mod_cluster does, but these calls appeared to work just fine outside of mod_cluster as it returns the following values from my test app:

12:08:26,519 INFO  [STDOUT] From MBeanServer: 1.81640625


So the issue does not appear to be necessarily with the underlying JDK/MXBean call but with how mod_cluster is handling the data grabbed from it.  But the root cause here appears to be that a value above 1 is being returned, and it looks like mod_cluster is expecting metrics to return a 0-1 percentile based range.

The way the load is determined allows the AverageSystemLoadMetric to improperly exceed its weight. For example if it were weighted as 2 and another metric was weighted at 1 (say RequestCountLoadMetric with a capacity of 1000), then AverageSystemLoadMetric should only be able to account for 67% of the load.  But here we can see AverageSystemLoadMetric can out run its weight and really account for 100% of the load.  So let's say AverageSystemLoadMetric is the above seen 1.81640625 value with 100 requests/second, putting RequestCountLoadMetric at .1 load, so DynamicLoadBalanceFactorProvider would calculate the load factor like so:

        int load = (int) Math.round(100 * totalWeightedLoad / totalWeight);
        int load = (int) Math.round(100 * (1.81640625 * 2 + 0.1) / 3);
        load = 124.4

And that gets truncated down to 100 so AverageSystemLoadMetric comes to represent really all of the load.  But if a 2 weight metric is at its max and a 1 weight metric is at .1%, then their total load should just be ~70-71%.

S is mod_cluster assuming that the SystemLoadAverage will always be between 0 and 1?  Does it look like mod_cluster is not properly scaling this metric?  Do we know a definite max return to expect from OperatingSystemMXBean.getSystemLoadAverage() so that this metric can be scaled more in line with the other percentile based ones?  Or should a user definable max capacity be implemented into this metric as it is with others?


--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.jboss.org/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira