high performance computing on graphics processing units: hgpu.org

Recent NVIDIA Graphics Processing Units (GPUs) can execute multiple kernels concurrently. On these GPUs, the thread block scheduler (TBS) uses the FIFO policy to schedule their thread blocks. We show that FIFO leaves performance to chance, resulting in significant loss of performance and fairness. To improve performance and fairness, we propose use of the preemptive Shortest Remaining Time First (SRTF) policy instead. Although SRTF requires an estimate of runtime of GPU kernels, we show that such an estimate of the runtime can be easily obtained using online profiling and exploiting a simple observation on GPU kernels’ grid structure. Specifically, we propose a novel Structural Runtime Predictor. Using a simple Staircase model of GPU kernel execution, we show that the runtime of a kernel can be predicted by profiling only the first few thread blocks. We evaluate an online predictor based on this model on benchmarks from ERCBench, and find that it can estimate the actual runtime reasonably well after the execution of only a single thread block. Next, we design a thread block scheduler that is both concurrent kernel-aware and uses this predictor. We implement the SRTF policy and evaluate it on two-program workloads from ERCBench. SRTF improves STP by 1.18x and ANTT by 2.25x over FIFO. When compared to MPMax, a state-of-the-art resource allocation policy for concurrent kernels, SRTF improves STP by 1.16x and ANTT by 1.3x. To improve fairness, we also propose SRTF/Adaptive which controls resource usage of concurrently executing kernels to maximize fairness. SRTF/Adaptive improves STP by 1.12x, ANTT by 2.23x and Fairness by 2.95x compared to FIFO. Overall, our implementation of SRTF achieves system throughput to within 12.64% of Shortest Job First (SJF, an oracle optimal scheduling policy), bridging 49% of the gap between FIFO and SJF.