Where two variational runs spend their time
A 3D-Var and a 3D-FGAT analysis, same mesh, same observations, same minimizer, 128 MPI tasks each. The pages below compare their cost and rank the code scopes worth optimizing.
Run timeline
Phase boundaries come from the OOPS_STATS milestone markers, so these segments partition the wall clock without overlap.
Resources
Wall clock, allocated core time, and resident memory as reported at the Run start and Run end markers.
Resident memory per task through the run
Traced from the per-milestone Local Memory field. The staircase inside the minimizer is the in-core Lanczos basis.
Hot spots
Every scope in the parallel timing table, ranked. Select a row for the full per-task spread and the other run's value.
Code map
Time attributed to the code that owns it. Select a wedge or tile to descend, use the trail above the chart to come back up.
Load imbalance
Fastest to slowest task for the scopes that waste the most core time. The marker is the mean; the bar spans min to max.
How to read these numbers
Three things about the source data change what the charts can and cannot tell you.