8 Steps to 3.7 TFLOP/s on NVIDIA V100 GPU: Roofline Analysis and Other Tricks

Abstract

Performance optimization can be a daunting task especially as the hardware architecture becomes more and more complex. This paper takes a kernel from the Materials Science code BerkeleyGW, and demonstrates a few performance analysis and optimization techniques. Despite challenges such as high register usage, low occupancy, complex data access patterns, and the existence of several long-latency instructions, we have achieved 3.7 TFLOP/s of double-precision performance on an NVIDIA V100 GPU, with 8 optimization steps. This is 55% of the theoretical peak, 6.7 TFLOP/s, at nominal frequency 1312 MHz, and 70% of the more customized peak based on our 58% FMA ratio, 5.3 TFLOP/s. An array of techniques used to analyze this OpenACC kernel and optimize its performance are shown, including the use of hierarchical Roofline performance model and the performance tool Nsight Compute. This kernel exhibits computational characteristics that are commonly seen in many high-performance computing (HPC) applications, and are expected to be very helpful to a general audience of HPC developers and computational scientists, as they pursue more performance on NVIDIA GPUs.