Section: New Results

Scheduling heuristics for dense linear algebra

In the context of Suraj Kumar 's PhD thesis, we are studying the scheduling of the Cholesky factorization on heterogeneous systems.

We have started to introduce communication costs into the constraint programming. Since this increases resolution time a lot, we had to optimize the expression of the data transfers to simplify the resolution. We modified the StarPU runtime system to be able to inject not only a static schedule of tasks, but also a static schedule of data transfers. This allows to inject the schedule optimized by constraint programming into real executions.

We have also shown how static schedules and dynamic scheduling strategies compare on heterogeneous platforms, and notably in the context of varying task execution time can typically be a problem for static scheduling. Static schedules have actually proven to be robust against variation in execution time. We have also studied injecting static information into dynamic schedulers, which improves the resulting performance with little offline analysis.