Using the Roofline Model and Intel Advisor

Series: HPC Best Practices Webinars

In this webinar, we will begin by introducing the Roofline Model and its “Cache-Aware” variant. We will proceed with some general guidelines and historical approaches to Roofline-based program analysis. Next, we will provide a short discussion of how changes in data locality and arithmetic intensity of two canonical benchmarks visually manifest in the context of these two Roofline formulations. Subsequently, we will provide two demonstrations of using Intel Advisor and the Roofline model within Intel Advisor. The first demo will be primarily instructive on how to compile, benchmark, and use Advisor. The second demo will focus on using variants of a simple benchmark to highlight changes in the Roofline model as well as providing correlation to Advisor’s other capabilities. We will conclude with a few comments on future directions.


Presenter Bios

Sam Williams is a staff scientist in the Performance and Algorithms Research Group at the Lawrence Berkeley National Laboratory (LBNL). He received his Ph.D. in Computer Science from the University of California at Berkeley in 2008. His research interests include high-performance computing, auto-tuning, performance modeling, computer architecture, and hardware/software co-design. He is currently involved in two ECP projects: AMReX (Adaptive Mesh Refinement for Exascale) and YTune (a compiler-based approach to auto-tuning). Sam is the main developer of the Roofline model.

Tuomas Koskela received his Ph.D. in applied physics from Aalto University (Finland) in 2015. His thesis work was on Monte Carlo modelling of fast ion confinement in Tokamak fusion reactors under 3D magnetic perturbations. Currently, Tuomas is a postdoc in the Exascale Science Applications Program of the National Energy Scientific Computing Center. His research interests include high-performance computing, plasma physics, nuclear fusion, fast ion physics.