A hybrid shared memory heterogeneous execution platform for PCIe-based GPGPUs

The disparity between the CPU and GPU domains has forced the programmers to adhere to the traditional driver-based GPU programming approach. The negative implications of this approach are inter-domain data transfer overhead, host memory pressure and CPU underutilization. In this paper, we propose a novel hybrid shared memory-based execution approach to enhance the throughput of the General Purpose GPU(GPGPU) applications. To achive optimal GPU execution, we adopted a midway approach between the shared memory and traditional disjoint memory GPU programming approach. Our design involves OS enhancements and extensions to an OS-integrated open-source GPU driver(GDev) which together provide the GPU application a shared memory execution platform. Our design not only eliminates several drawbacks associated with the traditional GPU programming approach, but allows data-parallel execution across CPUs and GPU.