Opened 3 months ago

Closed 3 months ago

#2703 closed Defect (fixed)

properly detect differences between simulations in prtctl

Reported by: smasson Owned by:
Priority: low Milestone:
Component: IOM Version: trunk
Severity: minor Keywords:
Cc: Branch review:
MP ready?: Task progress:

Description

Context

run.stat, computed in stpctl, is used to check if 2 simulations are identical or not.
prtctl is used to identify which tendency term is the first to create a difference between 2 experiments.

Analysis

When the 2 simulations differ, run.stat will show a difference but this difference between the run.stat files may arrive a few time step after the 2 simulations diverge.
run.stat is checking the min/max of some values. Using the min/max is convenient as it does not depend on the MPI domain decomposition. However the difference between 2 simulation is usually appearing somewhere between the min/max values and take a few time step to propagate until the min/max values.
I would say that this problem is not critical as the first goal of run.stat is to detect a potential difference.

prtctl is using a local sum to avoid this problem. However when the error is very small and the domain on which we apply the sum is large, the error can be hidden by the rounding done on the sum. This is a real critical point as prtctl must find the first term of the equation which diverges

Recommendation

use 2*wp precision in prtctl to make sure that all differences are really detected. Performance is bad but we don't care in this case.

Commit History (1)

ChangesetAuthorTimeChangeLog
15057smasson2021-06-25T09:43:49+02:00

trunk: use 2*wp precision in prtctl, #2703

Change History (2)

comment:1 Changed 3 months ago by smasson

In 15057:

trunk: use 2*wp precision in prtctl, #2703

comment:2 Changed 3 months ago by smasson

  • Resolution set to fixed
  • Status changed from new to closed
Note: See TracTickets for help on using tickets.