[Starpu-devel] question on CPU/GPU tasks distribution

["philippe.helluy@unistra.fr" <philippe.helluy@unistra.fr>]

Hi,

we have written a Discontinous Galerkin code for solving Maxwell's equation.
We have codelets with CPU and OpenCL functions, we activated the perf. 
model.

On small test cases and dmdar scheduler StarPU does a perfect job: ideal 
scaling for CPU-only runs and more than ideal speedup with CPU+GPU runs.

Things are worse on bigger tests. The code is faster with GPU-only runs 
(STARPU_NCPU=0) than with CPU+GPU runs. It seems that the dmdar 
scheduler is not able to take the decision to put all the computations 
on the GPU only, even when this is the better choice (I tried to change 
the beta to beta=2, with no effect). In this particular case, the whole 
data could be put entirely in the GPU memory.

Do you have hints on this behavior ?

Of course, I can provide more details if you are interested.

Best regards,

Philippe.