On Mon, Mar 9, 2009 at 9:40 PM, Berkin Ozisikyilmaz
<boz283@xxxxxxxxxxxxxxxxxxxx> wrote:
> ... since the combination of input parameters
> that I feed the applications change and hence for each of those I would need
> to create a checkpoint and it wouldn’t be a feasible way to do.
>
... or you can pick a Simics version and stick to it. It is well known
that small variations in initial conditions can have huge impacts on
parallel workloads. If your workload sees a lot of variance (aside:
how much does does performance vary run to run with a different
RANDOM_SEED and with the same Simics version?), you will have to run
additional data points to reach an acceptable confidence margin.
Regards,
Dan
--
http://www.cs.wisc.edu/~gibson [esc]:wq!
|