Bernie Velivis wrote:
> People have reported hitting limits around 1,500 users per OpenSTA server.
> Of course, you can have any number of servers.
There is an OpenSTA bug that limit the number of virtual users to 1664. I have been able to achieve this number before, but only with a trivially simple script. I have gotten up to 4800 virtual users total with 1600 users distributed across three load generators, and also with 1200 users spread across four load generators. I'm pretty sure I could push it higher than that by adding load generators or simplifying the script.
> Other factors can conspire to
> lower this number. If your scripts do a lot of string manipulation then CPU
> utilization on the OpenSTA server can limit the number of VUs. Very low
> think times will do the same.
I have a throughput test with no think time that can only get up to 25 virtual users before maxing out a single-cpu load generator. With a reasonable amount of think time, I can get about 10 times that many. That particular script doesn't do much string manipulation, but it has to download a large number of gif files, js files, css files, etc. for each primary page, which seems to load down the cpu heavily.
> On a dual processor 2.8Ghz P4 server running
> scripts that do a fair amout of string manipulation using think times of
> about 30 seconds, I usually plan for 500 users max until I see the
> application and get a feel for how complicated the scripts are.
That sounds like a decent rule of thumb. BTW, what measures do you use to determine that a load generator is maxed out and can't add any more virtual users without perturbing the performance numbers? I generally look for the cpu getting above about 80% utilization, which is usually the primary limiting factor in my experience.
When I'm running a stress test, though, and care more about beating up a server as much as possible without needing the performance timings, I'll push a load generator farther than that.
Danny R. Faught
Tejas Software Consulting