Resource Centered Computing Delivering High Parallel Performance

Modern parallel programming requires a combination of different paradigms, expertise and tuning, that correspond to the different levels in today's hierarchical architectures. To cope with the inherent difficulty, ORWL (ordered read-write locks) presents a new paradigm and toolbox centered around local or remote resources, such as data, processors or accelerators. ORWL programmers describe their computation in terms of access to these resources during critical sections. Exclusive or shared access to the resources is granted through FIFOs and with read-write semantic. ORWL partially replaces a classical runtime and offers a new API for resource centric parallel programming. We successfully ran an ORWL benchmark application on different parallel architectures (a multicore CPU cluster, a NUMA machine, a CPU+GPU cluster). When processing large data we achieved scalability and performance similar to a reference code built on top of MPI+OpenMP+CUDA. The integration of optimized kernels of scientific computing libraries (ATLAS and cuBLAS) has been almost effortless, and we were able to increase performance using both CPU and GPU cores on our hybrid hierarchical cluster simultaneously. We aim to make ORWL a new easy-to-use and efficient programming model and toolbox for parallel developers.