In 24/03/2008, infernix <[EMAIL PROTECTED]> wrote: > Hi, > > I've been playing with getting useful results from a very large dataset, > and am wondering whether if others have suggestions or better solutions. > > Jmeter ran with a testplan that has: > > - 1 thread group, 1200 threads, ramp-up 60s, scheduled for 6hrs
I would probably use a longer ramp-up for that many threads, especially since the test runs for a long time. > - 7 http requests, all with a response assertion (validates content), > and a Result Status Action Handler (stops thread) > - 1 Assertion Results that logs errors only > - 1 Aggregate Report that writes to the JTL file. Summary report uses less memory. [Simple data writer even less] Use CSV format for smaller output files, and only save the minimum fields needed. > - the point being: simulate a real-world test for a flash app that > queries a cluster with 9 requests per view, 20 viewers (aka clients) per > second, for 6 hours straight > > After letting this run for 6 hours, I got a 1.6GB large JTL file. I > already figured out that it's impossible to load into jmeter. It's also > impossible to parse with the XSL templates. I found out that xsltproc > and xalan are way too slow (took >10 minutes on a 33MB jtl file), but > after a long search I grabbed Saxon 6.5.5 > (http://saxon.sourceforge.net/) and this gave me good results. > > Saxon ran out of memory on the 1.5GB JTL file as well (with Xmx3000m, > sun-java6). It turned out that it could handle files of about 200MB. So > I hacked up a fairly simple piece of perl I found on the net to split it > in files of 750000 records each (http://pastebin.ca/955380). > > The resulting files can be processed by JMeter again, but the graphs > that are generated make no sense. So I have been playing around with the > xsl files in extras, and changed the jmeter-results-report_21.xsl > report a bit to include some extra information such as the time of the > first sample, last sample (e.g. test start+end times), and total MBytes > of data. This updated xsl (http://pastebin.ca/955390) does need > xsl-date-time (http://www.djkaty.com/drupal/xsl-date-time) to convert > unixtime into ISO time. > > So, I now have a HTML report for each 750000 records, and I simply > concatenate the 9 html reports into one HTML file. But that's not the > only information I need - I can't really make a useful graph out of it > that has a timescale. > > Next I used jtl2csv.py off the wiki and hacked it so it converts the > original 1.5GB Assertion Results JTL file into a Simple Data Writer log > (http://pastebin.ca/955398). I then hacked jtlmin.sh so it also adds > maxresponsetime in its OUT file (http://pastebin.ca/955409). So now I > have an OUT file with which I can actually create a graph per minute - > and this does work on the full test results, which is a good thing. > > After this learning curve, I stumbled upon StatAggVisualizer > (http://rubenlaguna.com/wp/better-jmeter-graphs/) and obviously I > immediately wanted to try it on my dataset. Unfortunately, it didn't > include makeTitlePanel() in the binary, so I grabbed the source, > uncommented that line, built it with ant and loaded it into JMeter. But > when I choose one of the split JTL files in the StatAggVisualiser > TitlePanel file field, it does not parse it and generate the graph. This > is rather unfortunate, as it seems that this one only works if you add > it to the test plan while it runs. > > Okay, so now for the questions part. > > - Am I doing things the right way [tm] with regards to the test plan setup? See above. > - How can JMeter itself generate useful results (read: graphs with a > timeline) out of very large tests like this? > > - Can someone look at the StatAggVisualizer and patch it so it loads > .jtl datasets and generate a graph out of it? I've tried this, but I am > simply not a coder. > > - Last but not least, if others have experience with testing large > datasets like this, and are willing to share their insights, I would > really appreciate it. There's stuff on the wiki but not all of that can > be applied to huge datasets. CSV files can be used as input to all sorts of stats packages and/or spreadsheets which may offer more flexibilty. > I've watched the Google presentation by Goranka Bjedov > (http://video.google.com/videoplay?docid=-6891978643577501895) which was > enlightening, but unfortunately there's little detail on the actual data > processing after it. It seems to me that they convert all their data to > SQL, insert it into MySQL and query it for useful results, but this > would mean that one has to develop a frontend for querying all that data > - which is beyond my project scope. > > Thanks in advance for any help! > > Regards, > > infernix > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [EMAIL PROTECTED] > For additional commands, e-mail: [EMAIL PROTECTED] > > --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
