Identifying Focus-of-Analysis Regions in MPI-Traces Using Transfer-Efficiency Monitors

Trace visualisation is the only reliable (identify structures [1] ) way to identify the different execution phases (Focus-of-Analysis or FoA) in HPC applications. Identifying them enables inspection of the differently behaving areas separately afterwards. Real-world HPC applications run long, almost entirely consisting of calculation phases (applications’ main work), while the initialisation or I/O durations are comparatively shorter. However, trace files are recorded for a much shorter duration due to time and disk space constraints, resulting in over-representation of the initialisation or domain decomposition steps compared to production runs. Any specific FoA must be identified in the timeline before analysis. Trace size increases with recording duration and number of MPI ranks, often making visualisation impractical. Even after visualising, complex communication patterns are hard to decipher. The appropriate outlook of exascale-ready methodologies is to identify and analyse FoAs in large trace files easily.