Speaker
Description
The ALICE detector at the LHC processed data rates of up to 1.2 TB/s
during Run 3. Real-time reconstruction of this continuous data stream is
essential, as the raw detector output is too large to be stored
permanently. ALICE therefore operates a heterogeneous computing farm
with 2800 GPUs, on which the most computationally demanding
reconstruction algorithms are executed.
This talk presents the software architecture and optimization techniques
that enable GPU-accelerated reconstruction in ALICE. Particular
attention is given to the Time Projection Chamber, which dominates the
online GPU workload. A common, portable implementation supports
different GPU architectures and is used for both real-time processing
and later offline reconstruction. The same code can therefore run on
GPU-equipped resources of the distributed computing Grid, avoiding
separate online and offline implementations.
Recent developments include automated, hardware-aware parameter
optimization and the exploration of neural networks for online cluster
reconstruction. Their impact on computing and physics performance will
be discussed, together with the remaining challenges and prospects for
future ALICE data-taking periods.