Hi!
> 
> Just in case someone is interested: As a Proof-of-Concept I started 100 thousand processes on a big machine (72 cores). It worked!
> However starting those too more than 30 minutes, and top needs more than 30 minutes to refresh ist display. Still, interactive input via SSH works nice, but any file-system access seems quite slow (my test processes just use CPU; the do no t do any I/O).
>

Okay, well. So you have kind of N^2 task, new processes need time and
all the all processes are already eating CPU... You are getting 55
processes per second, and that's not exactly good result.

> Kernel messages while the processes were created:
> kernel: [65648.247688] perf: interrupt took too long (2516 > 2500), lowering kernel.perf_event_max_sample_rate to 79250
> kernel: [65997.263218] perf: interrupt took too long (3146 > 3145), lowering kernel.perf_event_max_sample_rate to 63500
> kernel: [66790.221057] perf: interrupt took too long (3938 > 3932), lowering kernel.perf_event_max_sample_rate to 50750
> kernel: [69884.371426] perf: interrupt took too long (4925 > 4922), lowering kernel.perf_event_max_sample_rate to 40500
>

I'm seething these even without trying to overload the machine.

Best regards,
									Pavel

-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html