It seems that there is difference in behavior of http connector between CR1 and CR3 which influences our performance tests. This difference causes that test with CR3 ends sooner with less amount of clients.

Our client communicates directly to application server with http connector. We do not use Apache server in our tests. We have to to compare our performance results with IBM WAS and we have to have same test setup for both application servers.

Test for CR1 ends with more than 900 clients and mean response time is around 2400ms (test is stopped if mean response time is higher than 3000ms). Test for CR3 ends with Connection reset exception (there is timeout 30s) but mean response time is only 50ms.

What is the algorithm for choosing waiting requests for processing? Is it FIFO? We have to investigate this issue.

I have prepared following jobs on Hudson. It is clean configuration (without tuning in configuration).
http://hudson.qa.jboss.com/hudson/view/EAP51-Perf/job/eap-51x-stress-daytrader-ScenarioServlet-single-cr1-200-threads/
EAP 5.1 CR1 with default thread pool size (200)
Test ends with 950 clients

http://hudson.qa.jboss.com/hudson/view/EAP51-Perf/job/eap-51x-stress-daytrader-ScenarioServlet-single-cr3-200-threads/
EAP 5.1 CR3 with default thread pool size (200)
Test ends with 250 clients

http://hudson.qa.jboss.com/hudson/view/EAP51-Perf/job/eap-51x-stress-daytrader-ScenarioServlet-single-cr3-500-threads/
EAP 5.1 CR3 with increased thread pool size (500)
Test ends with 600 clients