high performance computing on graphics processing units: hgpu.org

The performance of many GPU-based systems depends heavily on the effective bandwidth for transferring data between the processors. For realtime systems, the importance of data transfer rates may be even higher due to non-deterministic transfer times that limit the ability to satisfy response time requirements. We present a new method that allows real-time applications to make efficient use of the communication infrastructure in multi-GPU systems, while retaining the necessary execution time predictability. Our method is based on a new application interface for executing batch operations composed of multiple command streams that can be executed in parallel. The new interface provides the run-time with information it needs to optimize the communication and to reduce the execution time. The method is compliant with common scheduling algorithms, such as EDF and RM, as it provides accurate offline execution time prediction for jobs using their definition and system characteristics. Experiments with two multi-GPU systems show that our method achieves 7.9x shorter execution time than the bandwidth allocation method, and 39% higher image resolution than the time division method, for realistic applications.