ASHA: An adaptive shared-memory sharing architecture for multi-programmed GPUs.

Abstract Spatial multi-programming is one of the most efficient multi-programming methods on Graphics Processing Units (GPUs). This multi-programming scheme generates variety in resource requirements of stream multiprocessors (SMs) and creates opportunities for sharing unused portions of each SM resource with other SMs. Although this approach drastically improves GPU performance, in some cases it leads to performance degradation due to the shortage of allocated resource to each program. Considering shared-memory as one of the main bottlenecks of thread-level parallelism (TLP), in this paper, we propose an adaptive shared-memory sharing architecture, called ASHA. ASHA enhances spatial multi-programming performance and increases utilization of GPU resources. Experimental results demonstrate that ASHA improves speedup of a multi-programmed GPU by 17%–21%, on average, for 2- to 8-program execution scenarios, respectively.