Speaker
Description
LHCb's Run 3 trigger is a fully software-based, two-stage system operating at the full 30 MHz bunch-crossing rate, a first for a collider experiment. HLT1, built on the GPU-accelerated Allen framework, performs the real-time reconstruction and selection that reduces the input rate to a manageable level, while HLT2, built on Gaudi, performs the full offline-quality reconstruction on the surviving events. This heterogeneous CPU/GPU model has been a remarkable success: it delivers the physics performance LHCb needs and has proven the viability of GPU-accelerated real-time analysis in a running experiment.
As we look towards Run 4 and Run 5, the computing challenge grows substantially. The most promising path forward is to port more, and eventually most, of the HLT2 reconstruction to GPU, so that the full trigger chain can benefit from the same acceleration that made HLT1 possible. Realising this ambition requires more than porting individual algorithms: it requires a single, coherent framework in which GPU and CPU algorithms can be developed, composed, tested, and run together, rather than the several disconnected execution paths that exist today.
This talk will present LHCb's experience with the Run 3 heterogeneous trigger, the lessons learned from operating two separate frameworks, and the ongoing consolidation effort that unifies them into a single scheduler supporting mixed GPU/CPU sequences. We will discuss the design of this unified framework, the new mechanisms that allow GPU and CPU data to flow between algorithms, and how this positions LHCb to progressively move HLT2 reconstruction onto GPU for Run 4.
