Speaker
Description
ATLAS is one of the two general-purpose experiments at the Large Hadron
Collider (LHC), aiming to detect a wide variety of physics processes. Its
trigger system plays a key role in selecting the events that are detected,
filtering them down from the 40 MHz bunch crossing rate to the 1 kHz rate at
which they are committed to storage. The ATLAS trigger works in two stages,
Level- 1 and the High-Level Trigger (HLT), with the first being a
hardware-based coarse filtering applied using custom electronics and FPGAs, and
the second relying on offline-like algorithms implemented fully in software,
running on a farm of commodity CPUs. The LHC will undergo the High-Luminosity
Upgrade soon (scheduled to be finished by 2029), which represents an additional
challenge for the ATLAS trigger. The increased pile-up leads to events that are
typically more complex and thus more computationally demanding to reconstruct,
and a broad-ranging suite of upgrades to the ATLAS detector itself also
encompasses increasing the input and output rates of the High Level Trigger by
a factor of 10. As such, both the processing power required to handle a single
event and the overall number of events that will need to be processed will
increase, placing greater pressure on the trigger farm. One possibility of
answering these increased computational demands in a cost- and energy-effective
way is the use of hardware accelerators, in particular leveraging the massive
parallelism and general computational capabilities offered by GPUs for problems
that are suited to their mode of operation.
Among the algorithms being assessed for GPU acceleration, Topological
Clustering, the main and most computationally demanding stage of calorimeter
reconstruction, has reached the significant milestone of 100% agreement with
the CPU algorithm and maximum speed-ups in excess of a factor of 10. This is
achieved through a more GPU-friendly variant of the algorithm, dubbed
Topo-Automaton Clustering. A significant bottleneck remains in the time taken
to convert between the data representation used within the GPU and the
equivalent CPU data structures, which can be up to two thirds of the total
execution time of the algorithm. This contribution will describe the
development, optimization and integration of Topo- Automaton Clustering with
the ATLAS trigger, including the latest benchmarks and ongoing efforts to
develop an EDM framework that could allow for a general description of
GPU-friendly data structures in order to alleviate the main bottleneck.
