high performance computing on graphics processing units: hgpu.org

Heterogeneous parallel architectures combining conventional multicore CPUs with GPUs and other types of accelerators promise significant performance gains compared to homogeneous systems. However, exploiting the full potential of such systems is becoming more and more challenging often forcing programmers to combine different programming models and parallelization strategies. A promising approach to coping with the increased programming complexity is the use of parallel patterns for expressing certain types of computations at a high-level of abstraction while relying on the compiler and runtime system to map such patterns onto a heterogeneous system. In this paper we present an approach for automatic performance tuning of high-level pipeline patterns for heterogeneous parallel systems in the context of a task-parallel component-based programming model. Our automatic performance tuning approach attempts to automatically determine the best combination of pattern-specific parameters, parameters exposed by the runtime system, and machine-specific parameters such that execution is optimized for a given workload and target architecture. Experimental results on two state-of-the-art heterogeneous systems demonstrate the effectiveness of our approach.