high performance computing on graphics processing units: hgpu.org

Experts agree that the exascale machine will comprise processors that contain many cores, which in turn will necessitate a much higher degree of concurrency. Software will require a minimum of a 1,000 times more concurrency. Most parallel analysis and visualization algorithms today work by partitioning data and running mostly serial algorithms concurrently on each data partition. Although this approach lends itself well to the concurrency of current high-performance computing, it does not exhibit the appropriate pervasive parallelism required for exascale computing. The data partitions are too small and the overhead of the threads is too large to make effective use of all the cores in an extreme-scale machine. This paper introduces a new visualization framework designed to exhibit the pervasive parallelism necessary for extreme scale machines. We demonstrate the use of this system on a GPU processor, which we feel is the best analog to an exascale node that we have available today.