1 of 1

Performance portability solutions for GPUs and CPUs to track reconstruction kernels�with the SciDAC-HEP NeuCol partnership

Scientific Achievement

We ported two mini-apps (p2z and p2r) from the mkFit application in the Compact Muon Solenoid (CMS) software framework to different CPUs and GPUs using various performance portability solutions and found the best mapping strategies to achieve the performance portability across diversely heterogeneous target systems.

Significance and Impact

High Energy Physics (HEP) experiments confront computing challenges from increasing detector sizes and accelerator intensities. The Compact Muon Solenoid (CMS) detector at the CERN LHC will face a factor of ~20 increase in reconstruction CPU time from the High-Luminosity LHC (HL-LHC) upgrade. Therefore, developing portable accelerator versions of the HEP application will be able to accelerate the software-based High-Level Trigger (HLT) across heterogeneous systems.

Summary of hardware supports for different portability solutions. Green indicates officially supported, red indicates unsupported.

We applied these portability solutions on two mini-apps (p2z and p2r) extracted from the HEP application called mkFit and evaluated these portability solutions, in terms of the development efforts and their performance on GPUs and many-core CPUs from different vendors (Intel, AMD, NVIDIA).

Technical Approach

Ported two important mini-apps (p2z) and p2r) to CPUs and GPUs using various programming models (Kokkos, SYCL, C++17 std::execution::par, Alpaka, OpenMP, and OpenACC).
Studied the performance of these mini-apps on the tested CPUs and GPUs and found the best parallelism and data mapping strategies.

PI(s): Robert Ross (ANL); Local Lab POC: Seyong Lee (ORNL)

Collaborating Institutions: ORNL, FNAL, United States Naval Academy, Cornell University, University of Oregon

ASCR Program: SciDAC RAPIDS2 ASCR PM: Kalyan Perumalla

Publication for this work: K. H. M. Kwok, et al., “Application of performance portability solutions for GPUs and many-core CPUs to track reconstruction kernels”, International Conference on Computing in High Energy & Nuclear Physics (CHEP), 2023.

LOCAL LAB POC: Seyong Lee (ORNL)

TALKING POINTS:

High Energy Physics (HEP) experiments are facing computing challenges from increasing detector sizes and accelerator intensities.
The Compact Muon Solenoid (CMS) detector at the CERN LHC will face a factor of ~20 increase in reconstruction CPU time from the High-Luminosity LHC (HL-LHC) upgrade.
Therefore, porting to the GPU-based heterogeneous system will be able to accelerate the software-based High-Level Trigger (HLT).
We port two mini-apps (p2z and p2r) from the mkFit application in the Compact Muon Solenoid (CMS) software framework to different CPUs and GPUs using various programming models such as Kokkos, SYCL, C++17 std::execution::par, Alpaka, and OpenMP/OpenACC, and find the best mapping strategies to achieve the performance portability across diversely heterogeneous systems.

METADATA:

Project: SciDAC RAPIDS2

PI name(s): Rob Ross (ANL), Scott Klasky (ORNL)
Name of the program manager: Supporting technology supported by ASCR research (Kalyan Perumalla)

Project: SciDAC-HEP NeuCol

PI name: Noemi Rocco (FNAL)
Name of the program manager: Supporting technology supported by ASCR research (Lali Chatterjee )

CITATIONS:

Ka Hei Martin Kwok, Matti Kortelainen, Giuseppe Cerati, Alexei Strelchenko, Oliver Gutsche, Allison Reinsvold Hall, Steve Lantz, Michael Reid, Daniel Riley, Sophie Berkman, Seyong Lee, Hammad Ather, Boyana Norris, and Cong Wang, Application of performance portability solutions for GPUs and many-core CPUs to track reconstruction kernels, International Conference on Computing in High Energy & Nuclear Physics (CHEP), 2023.