Re: [Starpu-devel] StarPU

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Hello,

Samy Chbinou, le Wed 25 Feb 2015 11:13:04 +0100, a écrit :
> 1/ Is it possible to choose the backend device that will do the work? ex 
> CPU or
> GPU or both in parallel.

You can specify the STARPU_NCUDA=0 environment variable to prevent CUDA
workers from starting, and the same with CPUS using STARPU_NCPU=0

> 2/ In case it is in parallel, is it possible to trace (I tried with FxT 
> wihtout
> success) the load for each Compute Unit?

You can use the STARPU_WORKER_STATS=1 environment variable to get a
summary, but FxT should have worked, please tell us what went wrong and
how.

> that is load in Mflops % on both CPU and GPU?

We haven't implemented counting Mflops for starpufft, that'd just be a
number of tasks.

> 3/ Can I integrate PAPI (for performance measurements)?

That has been on the TODO list for a long time, but it should feasible
to tinker something easily by adding PAPI initialization/destruction to
_starpu_cpu_worker().

Samuel