Greedily using GPU capacity for data list processing in multicore-GPU platforms

We have designed data list processing for multicore-GPU platforms and significantly improved the performance of both numerical and symbolic applications. For the latter, a novel aspect of our design was the management and processing of new data dynamically generated within GPUs. This paper presents various optimisations to our first design [1] aimed to use more the GPU, through reducing communication between the host (a multicore) and the GPU, in order to improve performance further. We present experimental results for three applications with different granularities and access patterns. Performance was improved again, significantly in some cases; using multicore-GPU platforms efficiently may involve complex changes to software.