Performance Engineering in High Energy Physics Software: The ATLAS Offline $\chi^2$ Track Fitter

As the ATLAS high energy physics experiment gears up for the HL-LHC upgrade, the performance of the software used to process the experiment's massive amount of data grows more and more important. Models predict that, when HL-LHC becomes operational in 2026, the amount of CPU time required to process the data produced by the ATLAS experiment will increase by up to a factor of nine; this increase exceeds by far the predicted computing budget. In this thesis, we present the results of a research project investigating the feasibility and applicability of performance engineering techniques on HEP software, using the ATLAS offline track fitting software as a guiding example. We discuss the challenges involved in optimising such software from a software and performance engineering perspective. Of particular interest are ways in which traditional performance analysis techniques can be adapted to obtain detailed and meaningful performance data for such software. Indeed, we find that not all techniques can be applied naively to give maximally meaningful results. From the results of performance analysis using techniques modified to fit this particular use case, we design and implement a series of relatively simple optimising code transformations. These optimisations approximately double the performance of the fitting code, thus bringing it closer to the performance required for HL-LHC