[Starpu-devel] What would you like to see most in the schedulers ?

[Cyril Roelandt <cyril.roelandt@inria.fr>]

Hi everyone!

We have just merged heft and dmda. What should we do next about our 
schedulers ? Here are some ideas:

* Merge pheft into dm*: I think Andra would like that to happen. I am 
wondering how hard this would really be, and if it can be done nicely.

* Make dm use compute_all_performance_predictions() in _dm_push_task(). 
It should not be too hard, but compute_all_performance_predictions() 
does a bit too much (like taking into account the power consumption). 
This would add bundle support to the dm strategy and would probably be 
helpful with regards to maintainability.

* Try to optimize dmda: this is probably still possible. For instance, 
this is probably quite bad:

    for (worker = 0; worker < nworkers; worker++)
    {
        ...
        for (nimpl = 0; nimpl < STARPU_MAXIMPLEMENTATIONS; nimpl++)
        {
            if (!starpu_worker_can_execute_task(worker, task, nimpl))
            {
                /* no one on that queue may execute this task */
                continue;
            }
            ...
        }
    }

Most codelets use only one implementation, so we'll end up calling 
starpu_worker_can_execute_task() (STARPU_MAXIMPLEMENTATIONS - 1) times 
for nothing. We could probably compute the exact number of 
implementations before entering these nested loops, and avoid a lot of 
function calls.

* Improve the readability and maintainability of dm*.

* Anything else ?


So, what would you be interested in ?


Cyril Roelandt.