Using Perftools and Reveal to Convert Applications to Run on GPUs, September 22, 2022

Introduction

This online training event hosted by NERSC will be presented by Senior Distinguished Technologist John Levesque from HPE, and open to NERSC, OLCF, and ALCF users. Perftools (Performance Measurement and Analysis Tools) and Reveal are available to collect CPU and GPU Performance on HPE systems, such as NERSC Perlmutter and ALCF Polaris with Nvidia A100 GPUs, and OLCF Frontier with AMD M250 GPUs.  

Date and Time: 9:00 am - 11:00 am (Pacific time), September 22 (Thursday) 

Overview

Within the next year, all DOE Office of Science centers will be offering a GPU-enhanced system for their users. Porting applications to run efficiently on a GPU is not an easy task and it is important to have a structured approach that gets the best possible performance. The preferred way to proceed is to use either OpenMP offload or OpenACC to get a majority of the application running on the GPU and then improve the inefficient bits on the GPU with CUDA or HIP. Hopefully, the entire application can achieve good performance without going to the low-level languages.

This talk will illustrate how to take an all MPI application and first use Perftools to identify which parts of the application should run on the GPU and then how to use Reveal to aid in adding OpenMP Offload directives to the original application. The talk will continue with showing how to gather runtime statistics and improve the performance of the applications.  

Registration

Please use this form to register.  This event will be presented online only using Zoom. 

Presentation Materials

Slides below.  Video Recording