Section: New Results

Simulation Execution with StarPU and SimGrid

The combination of StarPU and SimGrid allows to fast, accurate, and reproducible simulation the execution of task-based HPC applications.

This has proved to be very useful for theoretical research on scheduling heuristics [10] . It notably allowed to modify the simulated platform in order to easily observe and understand which parts of the platform (bandwidth, computation power) cause a bottleneck. It also allowed to remove some parts of the problem, such as the cost of data transfers, to simplify the problem and be able to deeply study scheduling solutions and compare them with optimum solutions in a simple environment before tackling the complete platform.

We have also extended the modelization of computation nodes, to take into account the PCI hierarchy of the system. This allows to get a more accurate simulation of systems which have dedicated channels between GPUs.

Last but not least, we have started to extend the StarPU+Simgrid combination to StarPU+MPI+Simgrid, to simulate the execution of HPC applications on clusters of heterogeneous systems. The preliminary results seem to show good accuracy. This will allow to easily study how applications scale, and study for instance how network performance have impact on it.