Re: [Starpu-devel] About StarPU

[Jeff DiMarco <jdimarco218@gmail.com>]

Thanks for the fast replay and answers!  The execution trace of the processing units will be very useful.  I have one more question.  If I have some application that can run on different resources, could I analyze the performance of this application after it has been mapped to these resources (possibly different configurations, too) ?

Thanks again,
-Jeff

On Mon, Oct 15, 2012 at 11:22 AM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Hello,

Jeff DiMarco, le Mon 15 Oct 2012 10:58:53 -0400, a écrit :

> I have just stumbled upon StarPU as a user-level scheduler for heterogeneous
> applications.  Is there any way for it to analyze performance of this
> scheduling easily?

Well, that depends on the application and what you want to measure. The
overhead itself of the scheduler can easily be evaluated thanks to the
tasks_size_overhead benchmark I have added to the trunk the other day.

The effectiveness of the heterogeneous scheduling is however harder to
evaluate. We usually simply measure the achieved application GFlops. You
can also measure how much data was transferred to evaluate how that goes
down with good anticipation by the scheduler.

You can also check with the execution trace whether processing units
were indeed busy, or more easily with both
STARPU_WORKER_STATS=1 STARPU_PROFILING=1

You can also check whether you are far from the theoretical upper bound,
see the corresponding chapter in the documentation.

Samuel