The O2 software framework and GPU usage in ALICE online and offline reconstruction in Run 3
- Eulisse, Giulio et al
- arXiv:2402.01205
 | Overview of the ALICE Experiment and its detectors in LHC Run 3 |
 | Illustration of the ALICE Run 3 synchronous reconstruction workflow for 8 GPUs and 128 virtual CPU cores. The workflow is split in the two NUMA domains, sharing only the shared input buffer. Both NUMA domains have 4 GPUs, which are each driven by individual OS processes. For simplicity, only the reconstruction processes are shown, and QC and calibration processes are omitted. |
 | Illustration of the processing graph of synchronous and asynchronous GPU processing steps in the baseline and optimistic scenarios. Colors indicate the readiness to run this step on the GPU. All steps are fully implemented and commissioned to run on CPU in the asynchronous processing. |
 | Speedup of several GPU models compared to one AMD Rome CPU core in the EPN servers. The measurements are corrected for the number of CPU cores required to drive a GPU, i.\,e.~the figure states how many CPU cores can be replaced by one GPU. |