Speaker
Description
On behalf of the ALICE collaboration
The ALICE experiment has a long-standing expertise in GPU computing for high-energy physics, having used hardware accelerators since Run 2. With Run 3, ALICE changed its data acquisition model to triggerless mode, relying even more heavily on GPUs to handle the data throughput produced by continuous detector readout. During Pb-Pb collisions today, Alice generates up to 3.5 TB/s of raw data, which must be stored in compressed form. The experiment's solution is to employ a two-stage compression chain during data taking, deployed on two separate computing farms. After a first low-level compression performed by FPGAs in the first farm, the raw data are reconstructed and compressed on the fly using 2800 GPUs, enabling the storage of the experiment data in a temporary disk buffer. After the online processing, the compressed data are subsequently reprocessed on the Worldwide LHC Computing Grid (WLCG) and on the online farm using GPUs during periods without beam, to produce physics data for analysis, in the so-called "offline" pass.
Processing the offline phase on the GPU-enabled online farm has already demonstrated a speedup of 2.5 when using the portion of reconstruction algorithms available on GPUs. Following these results, ALICE plans to port and optimize more algorithms to GPU, reaching an optimistic scenario where up to 80% of the current offline CPU time could be offloaded to GPUs. In parallel, ALICE is working to enhance the availability and monitoring of GPU resources on the grid, to further improve reconstruction performance for the offline pass.
This contribution will present ALICE’s experience with GPU computing, describing the GPU-powered data acquisition system, ALICE's vendor-agnostic GPU framework, and the performance improvements relative to CPU computing. It will also discuss future directions of GPU usage in ALICE, including the porting and optimization of reconstruction algorithms and the execution of GPU workloads on the grid.