Thanks for replying Philippe! I noticed that the test cases you guys used to test the performance of Jmeter is far to small than real world cases. Your test case had JUST 3 HTTP samplers with just 2 N-V pairs in each sample.
Real world scenarios have at least 100-200 HTTP samples per thread group. And the requests should contain at least 10-20 NV (name/value) pairs for POST requests and the responses should at least weigh 80-120KB. At least 20 response assertions. At least and at least 10 post response regex evaluators. This is when you actually exercise the engine. I am currently using Jmeter 2.8 and JRE 1.6. I am not using third party plugins. I am not using XML/CSV output. I have removed all listeners. I have 250 HTTP samplers with at least 10-NV pairs in each request. Most requests are of type: POST. I have 35 assertions and 13 regex evaluators. I have 2 If controllers, 15 transaction controllers, 8 constant timers and 1 thread group. Hope this helps. I will raise a defect in bugzilla but I'll be surprised if it doesn't already have one. Thanks Chaitanya M Bhatt On Thu, Nov 29, 2012 at 12:25 PM, Philippe Mouawad < [email protected]> wrote: > Hello, > Which version of JMeter are you using ? > > Can you attach your test plan to a bugzilla or show a description of it ? > > As for what you are saying, I have already used jmeter with a machine > simular to the configuration you are describing and went up to 4000 threads > and memory was not the limit in my case, > so you should investigate your test plan first particularly if you are > talking about Perm Gen as it's not the kind of memory that is highly > consumed except if you are using Javascript for example. > > - What kind of Test Elements are you using ? > - Are you using third party plugins? > - Which listeners have you setup > - Are you using XML or CSV output > > > As for what you are saying regarding "*This is a frustrating problem which > has been overlooked for a long time.*", please see: > > - http://wiki.apache.org/jmeter/JMeterPerformance > > Regards > > Philippe > > On Thu, Nov 29, 2012 at 8:40 PM, chaitanya bhatt > <[email protected]>wrote: > > > Group, > > > > I have noticed that Jmeter uses unreasonable amount of memory per thread. > > Even if you strip off loggers/result tree etc. you would still see a huge > > consumption of memory. I monitored the Jmeter memory heap and tuned the > JVM > > as much as possible. In spite of this I noticed that the tenured > generation > > of the heap is always packed with objects of huge size. Lots of classes > are > > loaded dynamically which is causing high occupancy in perm gen. > > > > I am working on another home grown tool which uses HTTPClient library to > > generate load. Using this home grown tool I can generate 4000 instances > per > > machine (64 bit with 32Gb ram and 16 CPU cores). But Jmeter on the > contrary > > fails to scale to a mere 1/10th of the target user load. > > > > This is a frustrating problem which has been overlooked for a long time. > > > > Can something be done about this? > > > > Thanks > > Chaitanya M Bhatt > > http://performancecompetence.com > > > > > > -- > Cordialement. > Philippe Mouawad. >
