Thursday, 11 Jun 2026, 15:00 CET
Traces are essential for HPC root-cause analysis, but full tracing at scale incurs prohibitive overheads and data volume. Selective tracing is one approach to combat this issue, by restricting trace collection on a structural and temporal level. The current tool ecosystem, however, offers limited support in this regard.
In this webinar, we showcase recent work towards effective compiler-assisted selective tracing. This is centered around a two-stage analysis workflow that semi-automatically derives trace configurations guided by parallel efficiency metrics. In a first stage, analysis on a whole-program call graph, extracted during link-time and embedded in the binary, is used to generate a low-overhead profiling configuration. Efficiency metrics collected in the profiling step are then queried by the analyst to generate selective trace configurations targeting performance-critical code regions.
We provide a runtime system that applies our selective tracing model on top of established performance tools, using LLVM XRay for efficient dynamic instrumentation filtering and introducing a light-weight call graph validation mechanism to ensure completeness of the static analysis results. We present a case study applying this workflow to a 500 node execution of Vlasiator, achieving a 10x trace size reduction while maintaining critical insight.
About the Presenter
Sebastian Kreutzer is a PhD candidate at the Scientific Computing lab at the Technical University of Darmstadt .
His research revolves around compiler-based tooling for HPC performance analysis.