Optimization of the computing performance of the inner detector track seeding algorithm used by ATLAS for High Luminosity LHC data

The High-Luminosity upgrade of the Large Hadron Collider (HL-LHC) will increase the instantaneous luminosity and associated pile-up to levels that impose unprecedented performance requirements on all stages of the event reconstruction chain. Track seeding, which constructs the initial trajectory candidates supplied to subsequent track-finding algorithms, has received comparatively limited, systematic optimisation within the ACTS framework, despite constituting a significant fraction of the total reconstruction time.

In this thesis, a two-phase profiling strategy is employed to identify and mitigate performance bottlenecks in the ACTS seeding algorithm. The first phase consists of lightweight sampling-based profiling for hotspot identification, followed in the second phase by source-level instrumentation and annotation to enable detailed root-cause analysis. Based on these studies, five targeted optimisations are proposed and evaluated using wall-clock timing, Callgrind instruction-count measurements, Heaptrack-based memory profiling, and physics-performance validation on a simulated dataset at HL-LHC pile-up conditions.

Of the five proposed modifications, three satisfy both the runtime improvement and physics-preservation criteria. These are combined into a new seeding implementation, Seeding3, which achieves a wall-clock speedup of 11.6× over the Seeding2 baseline while preserving equivalent physics performance. This corresponds to an increase in single-thread throughput from 1.7 to 19.3 events per second. These results demonstrate that the seeding stage possesses substantial headroom for performance optimisation when guided by systematic profiling, and that the adopted methodology is readily generalisable to other components of the reconstruction chain.