Hybrid static-dynamic selection of implementation alternatives in heterogeneous environments

With the emergence of heterogeneous architectures, developing parallel software has become an increasingly complex task. The ability of using multiple devices in a single application, such as CPUs, accelerators, or coprocessors, has turned the implementation and optimization tasks into a challenging process, which comes along with a variety of difficulties. The inherent complexities of the parallel algorithm, its multiple implementations, and the mapping possibilities onto one of the available processors are just examples of how intricate these tasks can become. To alleviate these issues, this paper proposes a hybrid static&-dynamic selector to better exploit resources provided by heterogeneous systems. Specifically, this framework generates at compile time a decision tree based on historical information for selecting the implementation that performs best at run-time. To evaluate the benefits of this approach, we analyze the performance with two use cases: the general matrix&-matrix multiplication and an image processing medical application. The experimental results demonstrate that our proposed selector enhances performance and minimizes efforts needed to tune applications. We proved that our solution improves from 10 to 24% the overall application performance in comparison with other similar approach.

Hybrid static-dynamic selection of implementation alternatives in heterogeneous environments Articles