Hi everyone, Thanks for developing tomcat, and kudos for having comprehensive JMX instrumentation for the server. I think its the best out of the box instrumentation one could ask for, for doing performance modeling and application monitoring.
I was a bit surprised when I tried out the Semaphore Valve and would like to get feedback on what others think. I noticed that the valve limits requests per second, based on the 'concurrency' configuration for the valve. However, I am not sure if this implementation is ideal, and the meaning of concurrency for Tomcat needs to be considered differently. My assertion is based on my prior work on performance modeling of Tomcat with Dr. Neil Gunther ( http://www.perfdynamics.com/Bio/njg.html). In queuing theory, Little’s Law expresses the relation between the average queue length, residence time and arrival rate. The average queue length can also be considered as the load in the system or the concurrency. In case of Tomcat, Little’s Law can be observed for the threads in service stage, requests per second and response time as: N = X * R, where N is concurrency or the threads in service stage, X is requests per second and R is the response time. The concurrency of tomcat then, is the number of threads which are actively serving requests. With the Semaphore Valve enabled (at the host level), however, I saw the requests per second being constant at the concurrency setting of the valve and number of service threads changing with a consistent jagged pattern of peaks. For eg, on a test host with the valve’s concurrency setting of 50 rps, the threads in service state went up to 700 and this formed a jagged pattern of peaks at 700 threads at regular intervals, unlike normal operation of Tomcat, where the number of service threads usually has a smoother pattern. What do you guys think ? Here’s a post I authored on Dr. Gunther’s blog on modeling Tomcat’s performance * http://perfdynamics.blogspot.de/2016/08/pdq-as-performance- periscope_3.html Note - the conclusion in the post needs revision, but the relations between key metrics discussed are valid. Thanks, Mohit
