Skip to main content

Hitting the Roof(line): Should I Optimize for Compute, Memory, or Both?

Speaker: Dunni Aribuki, Intel Corporatio

Presentation

Seminar Slide Deck (PDF-2.4MB)

Abstract

Building and deploying applications that run efficiently on modern hardware is complicated. Even the most skillful developers often struggle to understand performance in relation to hardware limitations, especially memory bandwidth and computational peaks. This talk presents the Roofline Model that provides a visual representation of an application’s performance against hardware-imposed performance ceilings. Roofline analysis provides developers with an optimization strategy to squeeze maximum performance out of their code with minimal time or effort. We will begin by discussing how roofline analysis helps to guide code optimizations for both memory and compute. Then, we will show how memory-access analysis helps to eliminate issues that can cause slowdown of vectorized code. We will provide application-specific optimization recommendations. We will conclude with a demonstration of Intel Advisor’s Roofline feature using a real-world particle physics application.

Biography

Photo of Dunni Aribuki
Dunni Aribuki is a Technical Consulting Engineer in the Compute Performance Developer Products Group at Intel. Her focus has been performance analysis and optimizations during her Intel tenure. Prior to joining Intel, she was a compiler engineer at Texas Instruments. In that role, she implemented OpenMP support in the DSP compiler and provided compiler support for various TI processors. Aribuki received a Ph.D. degree from the University of Houston in 2013.