Hybrid, Scalable, Trace-Driven Performance Modeling of GPGPUs

In this paper, we present PPT-GPU, a scalable performance prediction toolkit for GPUs. PPT-GPU achieves scalability through a hybrid high-level modeling approach where some computations are extrapolated and multiple parts of the model are parallelized. The tool primary prediction models use pre-collected memory and instructions traces of the workloads to accurately capture the dynamic behavior of the kernels. PPT-GPU reports an extensive array of GPU performance metrics accurately while being easily extensible. We use a broad set of benchmarks to verify predictions accuracy. We compare the results against hardware metrics collected using vendor profiling tools and cycle-accurate simulators. The results show that the performance predictions are highly correlated to the actual hardware (MAPE: < 16% and Correlation: > 0.98). Moreover, PPT-GPU is orders of magnitude faster than cycle-accurate simulators. This comprehensiveness of the collected metrics can guide architects and developers to perform design space explorations. Moreover, the scalability of the tool enables conducting efficient and fast sensitivity analyses for performance-critical applications.