Please visit Jefferson Lab Event Policies and Guidance before planning your next event: https://www.jlab.org/conference_planning.

May 8 – 12, 2023
Norfolk Waterside Marriott
US/Eastern timezone

Optimising the configuration of the CMS GPU reconstruction

Not scheduled
1h
Hampton Roads Ballroom and Foyer Area (Norfolk Waterside Marriott)

Hampton Roads Ballroom and Foyer Area

Norfolk Waterside Marriott

235 East Main Street Norfolk, VA 23510
Poster Poster Poster Session

Speaker

Mohamed, Abdulla (University of Bahrain)

Description

During the long shutdown prior to LHC Run 3, the CMS reconstruction software (CMSSW) was upgraded to offload 40% of the High Level Trigger (HLT) processing to GPUs. This upgrade accelerated the reconstruction algorithms and improved the efficiency of the HLT farm, however it introduced new parameters to the system that had to be selected carefully to maximise performance. When offloading parallel computations to GPUs, the choice of the optimal grid size and launch parameters is essential to achieve good occupancy of the GPU, and the best overall performance. The optimal grid size and launch parameters may depend on the properties of the kernels, on the size and dimensionality of the input data, and on the actual hardware being used. This contribution will present an implementation of a tunable interface for the GPU kernels in CMSSW, an autotuner for the interface with multiple tuning strategies, and discuss the performance characteristics of the different GPUs used in this work. The preliminary results of autotuning the CMS Patatrack pixel track reconstruction algorithms show a 6% increase in the throughput compared to the hand tuned version.

Consider for long presentation No

Primary authors

Mohamed, Abdulla (University of Bahrain) Dr Bocci, Andrea (CERN) Dr Elmedany, Wael (University of Bahrain) Dr Al-Ammal, Hesham (University of Bahrain)

Presentation materials

Peer reviewing

Paper