Abstract:

As general purpose computing graphic units are widely used in high-performance computing, a new concurrent execution model is proposed, under which the current memory scheduling policy is unable to achieve maximum memory throughput. We characterize different memory access behaviors of applications in the concurrent kernel execution on a single GPU platform, analyze the unbalanced performance loss across them, and propose a behavior-aware memory scheduling policy for GPGPU applications. Different priority scheduling methods are employed to exploit the advantages of application types. Experimental results show a significant improvement on the unbalanced performance loss among different types of applications. Averaged memory system throughput and fairness across all benchmarks are improved by 9.7% and 15.0% respectively over the baseline architecture.

Key words: GPGPU, concurrent execution, behavior-aware, memory scheduling