Increasing the Efficiency of the DaCS Programming Model for Heterogeneous Systems

Efficient programming of hybrid systems is usually done with the use of new programming models. It creates a unique opportunity to increase the performance of scientific applications and is also especially interesting in the context of future exascale applications development where extreme number of MPI processes tend to be a limitation. Future scientific codes will make use of hierarchical parallel programming models with message passing techniques used between nodes and optimized computational kernels used within multicore, multithreaded or accelerator nodes. In this article we consider the x86 and PowerXCell8i heterogeneous environment introduced in the High Performance Computing (HPC) sites like Roadrunner [6] or Nautilus [5]. Programming techniques for this environment are usually based on the IBM Data Communication and Synchronization library (DaCS). We describe our effort to increase the hybrid efficiency of the DaCS library and show how it affects the performance of scientific computations based on FFT kernels. The results are very promising especially for computational models that involve large three dimensional Fourier transformations.