high performance computing on graphics processing units: hgpu.org

Heterogeneous platforms play an increasingly important role in modern computer systems. They combine high performance with low power consumption. From mobiles to supercomputers, we see an increasing number of computer systems that are heterogeneous. The most well-known heterogeneous system, CPU+GPU platforms have been widely used in recent years. As they become more mainstream, serving multiple tasks from multiple users is an emerging challenge. A good scheduler can greatly improve performance. However, indiscriminately allocating tasks based on availability leads to poor performance. As modern GPUs have a large number of hardware resources, most tasks cannot efficiently utilize all of them. Concurrent task execution on GPU is a promising solution, however, indiscriminately running tasks in parallel causes a slowdown. This thesis focuses on scheduling OpenCL kernels. A runtime framework is developed to determine where to schedule OpenCL kernels. It predicts the best-fit device by using a machine learning-based classifier, then schedules the kernels accordingly to either CPU or GPU. To improve GPU utilization, a kernel merging approach is proposed. Kernels are merged if their predicted co-execution can provide better performance than sequential execution. A machine learning based classifier is developed to find the best kernel pairs for co-execution on GPU. Finally, a runtime framework is developed to schedule kernels separately on either CPU or GPU, and run kernels in pairs if their co-execution can improve performance. The approaches developed in this thesis significantly improve system performance and outperform all existing techniques.