Contrast and Analysis about the Characteristics of MPS and CDP in GPU Kepler Architecture

The new generation architecture of NVIDIA launched Multi-Process Services (MPS), which provides a context manager in the software layer to handle tasks with different processes. MPS can only be used on the Linux platform, and requires a computing capability of 5.0 or higher NVIDIA GPU card [1]. Although these constraints limit the applicability, but it is a relatively inexpensive way to make multiple processes take full advantage of GPU resources. CUDA Parallel Dynamic (CDP) is the other new execution model introduced in Kepler GK110, which allows GPU kernel function to create additional task for itself. It can control the new task scheduling work and synchronization the results without CPU intervention [2]. So that CUDA application is no longer constrained by the rule that the kernel function on GPU must be called from the CPU program. Kernel function can call a new kernel function directly from the GPU kernel function, thereby enhancing the GPU parallelism.