high performance computing on graphics processing units: hgpu.org

The increasing diversity of hardware within a single system promises large performance gains but also poses a challenge for data management systems. Strategies for the efficient use of hardware with large performance differences are still lacking. For example, existing research on GPU supported data management largely handles the GPU in isolation from the system’s CPU | The GPU is considered the central processor and the CPU used only to mitigate the GPU’s weaknesses where necessary. To make efficient use of all available devices, we developed a processing strategy that lets unequal devices like GPU and CPU combine their strengths rather than work in isolation. To this end, we decompose relational data into individual bits and place the resulting partitions on the appropriate devices. Operations are processed in phases, each phase executed on one device. This way, we achieve significant performance gains and good load distribution among the available devices in a limited real-life use case. To grow this idea into a generic system, we identify challenges as well as potential hardware configurations and applications that can benefit from this approach.