[Starpu-devel] Question regarding StarPU's coherency model

[FLINT Clément (ETU MAI) <clement.flint@etu.unistra.fr>]

Greetings,

I am currently working with Bérenger BRAMAS (copy) on the heteroprio scheduler in starpu.

In order to implement good heuristics, we need to have access to tasks successors and predecessors.

When a ready task is pushed to the scheduler, we would want to compute a score based on this task's successors's predecessors (which may or may not be ready). Is there any standard way to get this information in starpu ?

So far we have tried 3 approaches which did not lead to a satisfying result:

- using handles' req_list fields (coherency.h ~ 118). This seems to be designed to keep track of data requests on a handle and we can theoretically deduce the information we need. Unfortunately, in our test cases, this list was always empty and therefore unusable

- using the _STARPU_JOB_GET_DEP_SLOTS macro. It is unclear for us what this macro should return. In our tests it returned the tasks themselves or an other task that is allowed to read concurrently

- using the post_sync_tasks field of the handle structure (coherency.h ~ 226). This field seems to contain successors to the current active task group of the handle. However, this would not be sufficient for our purpose, since we would like to get the predecessors of an unready task.

In view of this, we would have the following question: does starpu provide any straightforward way of browsing through successors and predecessors, or would heteroprio need to maintain its own request lists for each handle in order to do so ?

Cordially.

Clément FLINT