Dynamic Resource Scheduling on Graphics Processors

Graphics processors offer tremendous processing power, but do not deliver peak performance, if programs do not offer the ability to be parallelized into thousands of coherently executing threads of execution. This talk focuses on this issue, unlocking the gates of GPU execution for a new class of algorithms.

We present a new processing model enabling fast GPU execution. With our model, dynamic algorithms with various degrees of parallelism at any point during execution are scheduled to be executed efficiently. The core of our processing model is formed  by a versatile task scheduler, based on highly efficient queuing strategies. It combines work to be executed by single threads or groups of thread for efficient execution.

Furthermore, it allows different processes to use a single GPU concurrently, dividing the available processing time fairly between them. To assist highly parallel programs, we provide a memory allocator which can serve concurrent requests of tens of thousands of threads. To provide algorithms with the ultimate control over the execution, our execution model supports custom priorities, offering any possible scheduling policy. With this research, we provide the currently fastest queuing mechanisms for the GPU, the fastest dynamic memory allocator for massively parallel architectures, and the only autonomous GPU scheduling framework that can handle different granularities of parallelism efficiently. We show the advantages of our model in comparison to state-of-the-art algorithms in the field of rendering, visualization, and geometric modeling.