On 24/02/2011 06:42, Tony Anecito wrote: > Hi Mark, > > I setup a simple servlet with no code in the doGet() method running in Tomcat > 7.0.8 and using jvm 1.6.22 and using httpclient from apache commons lib and I > am > lucky to get 2.8msec per request as measured at the client. I am thinking the > commons httpclient is slow itself. > > So what did you use for the client side?
I was using JMeter. Key configuration was: - servlet that did nothing - http keepalive enabled - dual core test machine - single JMeter thread - all on localhost The above configuration meant that there was only a single thread used on Tomcat. The results on OSX were *much* better than those on Windows although I don't recall the details. > Also, I looked at the test results someone ran using Apache httpclient and > they > took the total number of requests and the time it took to run and came out > with > an absurd number because the response times were much longer than the number > they calaculated. I could do the same thing by say taking 100 requests that > ran > in parallel and took 100msec each and divide by 100 and say it took 1msec per > request but that is not what the user would see. I've done some tests along those lines in the past with the same simple servlet and the result of (total time)/(total number of requests) was largely similar regardless of how many requests were processed in parallel. In short, the more threads, the longer each request took. I saw the best throughput where parallel threads was at or just above the number of cores available (unsurprisingly). A real app is likely to be different. > Just interesting how people can twist statistics. Indeed. We often say on this list that you can construct a load-test to show that any combination of Tomcat connector with/without httpd is the 'best'. It all depends on the test. Mark --------------------------------------------------------------------- To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org For additional commands, e-mail: users-h...@tomcat.apache.org