Speaker: Laura Cappelli, CERN

To fully exploit the physics reach of the High-Luminosity Large Hadron Collider (HL-LHC), the LHC experiments are planning substantial upgrades of their detector technologies and increases of their data acquisition rates. The higher proton-proton interaction rate, pile-up, and event processing rates present an unprecedented challenge to the real-time and offline event reconstruction. This requires a processing power that is orders of magnitude larger than today, and exceeds by far the expected increase for conventional CPUs. The Compact Muon Solenoid (CMS) experiment is developing a fully heterogeneous reconstruction software that will be used during the next LHC data-taking period, starting in 2022. Its first applications will be the online reconstruction running on a GPU-equipped High-Level Trigger (HLT) farm, and the offline reconstruction running on HPC centers worldwide. These activities will allow the collaboration to gain experience with parallel algorithms and a heterogeneous framework that will be essential to leverage diverse kinds of accelerators during the HL-LHC data taking. To keep under control the cost of software development, maintenance, and validation that this will entail, CMS is evaluating various performance portability frameworks that promise a “write once, run anywhere” approach, building the same code base for different back-ends and accelerator types.

This talk details the ongoing work to port the CMS reconstruction software to the Intel® oneAPI platform and compare its performance on different back-ends with that of native code running on the same hardware.