Parallel Data List Processing on Multicore-GPU Platforms

Multicore-GPU platforms are now common and affordable, yet capitalising on their parallel processing capability is not straightforward. Existing sequential and parallel software must be tuned, or designed anew, to efficiently capitalise on these platforms. This paper presents the design of parallel data list processing in multicore-GPU platforms, wherein application data is organised into various lists, one list for each core and GPU device, for the purpose of balancing the workload through work (data items) stealing. A novel aspect of our design is the processing of new data dynamically generated within GPUs. We present experimental results for three applications with different granularities and access patterns. Overall the use of GPUs can significantly improve performance, but using them profitably may not be simple.