"Addressing the software and computing challenges of physics event generators"
The HL-LHC physics program will require unprecedented computing resources for simulated collisions and therefore Monte Carlo (MC) event generation. The number of MC events to generate and simulate, which scales as the integrated luminosity of real collisions, is expected to increase at a much faster rate than the available computing resources. The fraction of CPU consumed by MC event generators will also increase dramatically, as the LHC experiments are expected to use primarily fast detector simulation. In addition, the availability and reliance upon higher precision theoretical calculations is also expected to increase, resulting in increased computing resource requirements for MC generation compared to today. Consequently, to ensure physics results are not restricted by a limited number of MC events that can be generated on the available resources, significant optimisation upon the current model is required.
To achieve this, first a detailed analysis of the current computing model for MC event generators is required, followed by optimisation to improve generator software performance. In addition, developing generator software to be able to exploit the evolution of computing architectures in the HL-LHC era, which will likely take the form of accelerator-based High Performance Computing clusters, will be vital. Such architectures are better suited to the smaller code-bases of MC generators rather than the vast code bases of the LHC experiments simulation and reconstruction infrastructures. Finally, more precise projections of the generator-level physics requirements for HL-LHC are also required from theorists and experimentalists.
The HEP Software Foundation has recently created an Event Generators Working Group that aims to bring together MC generator authors, LHC experiment users and software experts to address these issues. The first results of this working group are presented here. This includes a detailed study of the ATLAS and CMS generator usage during Run 2, benchmarking of different MC generator CPU performance, and first studies on using MC generator code on new architectures.