Description
This work introduces an expanded event model in Vampir for analyzing non-blocking MPI communication.
This model addresses the challenges of unsuccessful tests and attributing wait time to root-cause functions in application code.
These extended message events allow Vampir to visualize temporal and spatial high-frequent test-phases and wait-time distributions, supporting outlier detection and function-level summaries that go beyond traditional time-based statistics.
We demonstrate this approach using the OpenLB’s rayleighTaylor3D benchmark and show that excessive MPI_Test calls do not degrade runtime and that reducing the test frequency induced due to the measurements overhead maintains the performance.
An extended Score-P recording mode captures unsuccessful test events with negligible size and time impact.
Our results confirm that applications do not need to aggressively optimize MPI_Test frequency and that Vampir can effectively identify wait-time root causes.