Task lead: Axel Naumann (CERN)

To make efficient use of accelerator (GPU and FPGA) devices in the software designed for the High-Luminosity LHC, various common developments and improvements are needed in the frameworks and code bases of the experiments and Monte Carlo generators. Frameworks need to make efficient use of all available computing resources of single compute nodes, and even possibly multiple nodes at the same time. Existing implementations should be harmonized between the experiments, and optimization efforts should be shared.

In order to make the software used by the experiments, TH and IT as efficient as possible, we need to develop techniques and improvements in the following areas:

● Efficient scheduling of computing steps on heterogeneous devices for ensuring that CPU and accelerator resources are maximally utilized;
● Efficient data structures for heterogeneous software to ensure that memory copies are minimized and are as efficient as possible;
● Common High-Energy Physics libraries for heterogeneous systems for implementing optimal calculation of common operations for HEP code, running on accelerator devices;
● Efficient interfaces to Machine Learning inference engines to minimize data movements and execution latencies;
● Finally, novel programming languages should be evaluated for the implementation of reconstruction algorithms in the High Level Triggers.