Accelerating End-to-End Deep Learning for Particle Reconstruction using CMS open data

Michael Andrews,Bjorn Burkle,Meenakshi Narain,Emanuele Usai,Manfred Paulini,Ulrich Heintz,Davide Di Croce,Sergei Gleyzer,Shravan Chaudhari,C Biscarat,B Hegner,G.A Stewart,S Roiser,S Campana,C.I Rovelli

doi:10.1051/epjconf/202125103057

Abstract

Machine learning algorithms are gaining ground in high energy physics for applications in particle and event identification, physics analysis, detector reconstruction, simulation and trigger. Currently, most data-analysis tasks at LHC experiments benefit from the use of machine learning. Incorporating these computational tools in the experimental framework presents new challenges. This paper reports on the implementation of the end-to-end deep learning with the CMS software framework and the scaling of the end-to-end deep learning with multiple GPUs. The end-to-end deep learning technique combines deep learning algorithms and low-level detector representation for particle and event identification. We demonstrate the end-to-end implementation on a top quark benchmark and perform studies with various hardware architectures including single and multiple GPUs and Google TPU.

Highlights

One of the major challenges for high-energy physicists is to manage the large amounts of data generated by the Large Hadron Collider (LHC)
Most machine learning-based particle identification techniques currently developed by ATLAS and CMS experiments rely on inputs provided by the Particle Flow (PF) algorithm used to convert detector level information to physics objects [1]
We demonstrate the integration of the end-to-end deep learning framework into CMS Software Framework (CMSSW) on a simulated top quark benchmark [6] and compare the performance of realistic deep learning algorithms on single and multiple Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs)

Summary

Introduction

One of the major challenges for high-energy physicists is to manage the large amounts of data generated by the Large Hadron Collider (LHC). To attain the main physics goals of the HL-LHC, the use of advanced machine learning techniques will be necessary to overcome significant challenges posed by increased levels of pile-up and the rarity of sought-after signals To address this issue, researchers are applying state-of-the-art machine learning algorithms for data processing and detector reconstruction, aimed at optimising their performance and accelerating these models during training and inference. The end-to-end deep learning technique [3, 4] combines deep learning algorithms and low-level detector representation of collision events The integration of such innovative machine learning algorithms with the data processing pipeline of experimental software frameworks is an important goal for LHC experiments [5] and forms one of the major research and development goals for CMS. We demonstrate the integration of the end-to-end deep learning framework into CMS Software Framework (CMSSW) on a simulated top quark benchmark [6] and compare the performance of realistic deep learning algorithms on single and multiple GPUs and TPUs

Open Data Simulated Samples

Integration of End-to-End Deep Learning with CMSSW Framework

Typical End-to-End Framework Pipeline

Scaling Deep Learning Training to Multiple GPUs

Timing Performance Comparison

Conclusions