Applying the Roofline model for Deep Learning performance optimizations

TL;DR

This paper introduces an automated methodology for constructing Roofline performance models tailored for NUMA architectures, exemplified with Intel Xeon, and evaluates optimized deep learning primitives from Intel oneDNN.

Contribution

The paper presents a novel automated approach for Roofline modeling on NUMA systems and assesses the performance of optimized deep learning primitives.

Findings

Automated Roofline models effectively represent NUMA system performance.

Optimized deep learning primitives show significant performance improvements.

The methodology facilitates performance analysis and optimization for complex architectures.

Abstract

In this paper We present a methodology for creating Roofline models automatically for Non-Unified Memory Access (NUMA) using Intel Xeon as an example. Finally, we present an evaluation of highly efficient deep learning primitives as implemented in the Intel oneDNN Library.