GeantV

Abstract

Full detector simulation was among the largest CPU consumers in all CERN experiment software stacks for the first two runs of the Large Hadron Collider. In the early 2010s, it was projected that simulation demands would scale linearly with increasing luminosity, with only partial compensation from increasing computing resources. The extension of fast simulation approaches to cover more use cases that represent a larger fraction of the simulation budget is only part of the solution, because of intrinsic precision limitations. The remainder corresponds to speeding up the simulation software by several factors, which is not achievable by just applying simple optimizations to the current code base. In this context, the GeantV R&D project was launched, aiming to redesign the legacy particle transport code in order to benefit from features of fine-grained parallelism, including vectorization and increased locality of both instruction and data. This paper provides an extensive presentation of the results and achievements of this R&D project, as well as the conclusions and lessons learned from the beta version prototype.

Detector Simulation Challenges for Future Accelerator Experiments

“…Successive upgrades to adapt to new computing paradigms such as object-oriented or parallel design have not touched the main modelling concepts described above, which served their purpose for decades of CPU evolution but are quickly becoming a limiting factor for computing hardware with acceleration. Recent R&D studies [27,28] have shown that today's state-of-the-art Geant-derived geometry codes such as VecGeom [29] represent a bottleneck for vectorized or massively parallel workflows. Deep polymorphic code stacks, low branch predictability, and incoherent memory access are some of the most important reasons for performance degradation when instruction execution coherence is a hardware constraint.…”

“…Moreover, new technologies [41][42][43] will allow detectors to sample particle showers with a high time resolution of the order of tens of picoseconds, which will need to be matched in simulation. Consequently, the simulation community has launched an ambitious R&D effort to upgrade physics models to improve accuracy and speed, re-implementing them from the ground up when necessary (e.g., GeantV [27], Adept [28], Celeritas [44]). Special attention will be needed to extend accurate physics simulation to the O (100)TeV domain, including new processes and models required to support the future collider programs.…”

“…Seeking to obtain massive parallelism of thousands or tens of thousands of active particles is challenging to develop in detector simulation. The GeantV project [27] explored the potential of SIMD-CPU based parallelism by marshalling similar work ('event-based' in the parlance of neutron simulation), e.g., waiting till many particles entered a particular volume before propagating the particles through that volume. The project's conclusion was that the speedup potential was modestbetween 1.2 and 2.0.…”