I'm having some difficulty testing different parallelization methods and parameters on my Xenode, in that test result variability is high.

Since the physical server is shared, any other CPU usage appears as a slower test in my results - sometimes taking 4 times as long as the minimum time. Re-running the same test repeatedly yields varying times, yet with a floor minimum time (presumably when the hardware CPUs are all available).


There seem to be two ways to reduce the testing uncertainty.

1) I repeat the tests many times, taking the minimum test time as the result. This slows down testing and takes away resources from other users, so I don't like this method at all. I can never be sure that I've hit a time when all CPUs were avaiable to me, and the only way to reduce this uncertainty is to repeat the tests more and more often for a greater statistical sampling.

2) Force my share of the Xenode CPU utilization to minimum. This reduces CPU availability uncertainty to zero (they will never run any slower than this), so I only run my tests once. This also has the absolute minimum possible impact on other users. So I like this method.

How can I force my Xenode to minimum CPU share? Is there a way?

James

I don't think it's possible; most OSes (including probably Xen) won't leave the CPU idle when there's a runnable domain.

Well your best solution is to use a good testing method, ala not run it on a shared machine.

Can I borrow one of your multi-core servers for testing?

James

Allow me to rephrase the question: Could it be set it to use *no more than* my minimum share?

James

I think amazon ec2 has hard limit on cpu resources. Their xlarge instance has 8 cores.