Section: New Results

Interfacing the MPC Parallel Framework with StarPU

CEA has developped a framework named MPC that transforms MPI+OpenMP applications into a lightweight thread-based program which can flexibly and efficiently exploit multicore architectures. StarPU, on its side, is mainly dedicated to schedule coarse grain tasks over accelerators, and is less suited to fine grain task scheduling. We have thus initiated a software interoperability effort between StarPU and MPC. The first step was to implement a new StarPU task scheduling strategy based on a NUMA-aware adaptative task granularity according to the target device (GPU or CPU). We observed significative performance gains for a Cholesky application in comparison to an eager strategy, thanks to the NUMA-aware aspect. However more work is still needed with respect to task decomposition as it implies data partitioning during the execution. We are also working on a variable granularity task programming interface in order to simplify the developer's coding effort. Finally, we develop a mechanism in StarPU to isolate some parts of the computing platform for another runtime. We used nested scheduling contexts to redirect some tasks to a scheduling component that StarPU may or may not control. The idea is to associate a scheduling subcontext to a runtime, for instance MPC, that would access to a dedicated set of computing resources for executing parallel kernels.