Using GPU to Reconstruct LHC Collisions Recorded with the CMS Detector

PI Dirk Hufnagel , Fermi National Accelerator Laboratory
Co-PI Stephan Lammel, Fermi National Accelerator Laboratory
Project Summary

This team set out to use GPU resources at ALCF including Polaris and Aurora to identify the fundamental constituents of matter, to precisely measure the forces between them, to identify new symmetries, and to search for dark matter.

Project Description

The Compact Muon Solenoid, CMS, detector records high-energy proton-proton collisions of the Large Hadron Collider, LHC. The CMS collaboration operates the detector and analyzes the collisions to search for the fundamental constituents of matter, to precisely measure the forces between them, to identify new symmetries, and search for dark matter. Position, timing, and energy deposits recorded by the detector are reconstructed back to particles generated in the collision and their momenta. These collisions are compared to collisions simulated with Monte Carlo techniques from the standard model, SM, extension of the SM, and other theoretical models to gain insight into nature at small scale or the conditions in our universe at about a picosecond age.

So far high-energy physics, HEP, experiments have used local and grid computing resources to reconstruct and simulate collisions. After the next upgrade of the LHC, the collisions will be so complex that grid resources alone will be insufficient. CMS are exploring high-performance computing, HPC, usage since a few years. HPC resources pose two challenges for CMS and HEP experiments in general:

1. The grid computing environment reflects the open collaborative approach in large HEP experiments. HPC resources, and especially leadership class facility HPC resources, have a tighter security model. This makes integration more difficult especially for our data intensive workflows where data needs to be transferred into and out of the site.

2. HEP software was written for CPUs. Migrating our software to use GPU started a few years ago. A number of algorithms have already been adapted and will be used in the processing of the upcoming data taking. For high-luminosity, HL-LHC, we anticipate simulation, reconstruction, and even some analysis workflows to fully harness the compute power provided by GPUs.

The objective of this proposal are two fold:

• to use GPU resources at ALCF Polaris/Aurora for the reconstruction of some of the Run3 data collected at the CMS detector at the LHC

• to develop and exercise better integration of HPC (LCF) resources into the CMS computing infrastructure.