Re: [Starpu-devel] StarPU code hangs when using a lot of GPU memory

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

On 2019-03-01 18:18, Samuel Thibault wrote:
> It actually is: nosubdataref and nodataref mean that all of the 5815 
> are
> considered busy, and that's the reason why they can't be evicted.
>
> Could it be that you have a *lot* of tasks which are ready? It might
> just be that they all almost have their data on the GPU, but none of
> them has all it needs, and thus none of them can be executed. It is a
> fairly extreme situation that we did envision, but didn't put 
> management
> in StarPU yet. That's however precisely the kind of work I'll be having
> a look at in the coming months :)

There is definitively a lot of tasks ready to be scheduled and these 
tasks would end up updating almost the entire matrix in small sections. 
The entire matrix cannot be fitted into the GPU memory so the situation 
you are describing is likely to happen with my code.

Is StarPU smart enough to move tasks from the CUDA worker to the CPU 
workers? The scheduling context should have a large pool of CPU workers 
at the point were the programs stalls.

 - Mirko