I am running Apache DayTrader web primitives tests on JBoss EAP 5.1 and when using natives the performance drops 3x.

To stress the web server the troughput is reaching about 17000 requests per seconds therefore I use a large HTTP connector thread pool (maxThreads is set to 6000).

I have noticed that with small load (200 clients, 100ms delay between requests) the results are very similar:

Sessions: 200,throughput 1,753.4 samples/s,mean response: 11 ms
Sessions: 200,throughput 1,760.9 samples/s,mean response: 10 ms

however when running at "full speed", the throughput drops from 17,507.2 samples/s to 5,832.8 samples/s which seems to be a problem with how natives is handling large thread pool sized.

I would be very thankful if we could figure this out, understand why does it drop and how it can be fixed or avoided, since our documentation doesn't say anything about configuration (is it even configurable?), just how to install it.

Some runs. With:
http://hudson.qa.jboss.com/hudson/view/EAP51-Perf/job/eap-510-load-primitive-PingJspEL-single/8/

Without:
http://hudson.qa.jboss.com/hudson/view/EAP51-Perf/job/eap-510-load-primitive-PingJspEL-single/9/

—

Related problem is the unfairness which affects both Java and native. In the test we have 1500 clients so 800 threads are handling 800 requests concurrently, so some have to be queued and since the connector is unfair some of the requests timeout - which is set to 30 seconds.

—

Apart from the original problem, I decreased maxThreads to 800 (4 cores x 200 threads). Now natives compared to Java show ~0.7% performance increase at 150 clients, at 1500 clients it ~0.5%.