Modify

Opened 6 years ago

Last modified 5 years ago

#1221 new defect

Reduction weight anomaly when using zero CoordBase::boundary_shiftout_*

Reported by: Eloisa Bentivegna Owned by: Erik Schnetter
Priority: minor Milestone:
Component: Carpet Version: development version
Keywords: Cc:

Description

When using a refinement level that covers the entire grid, and setting the CoordBase::boundary_shiftout_* parameters to zero, the following warning appears:

INFO (CarpetReduce): Simulation domain volume: 8000
INFO (CarpetReduce): Additional excised volume: 0
INFO (CarpetReduce): Reduction weight sum: 9141
WARNING level 1 in thorn CarpetReduce processor 0 host socket.local

(line 137 of /Users/bennie/CactusTrees/TrueAMR/arrangements/Carpet/CarpetReduce/src/mask_test.c):
-> Simulation domain volume and reduction weight sum differ

This can be reproduced using the attached parameter file. The problem seems to arise at the boundaries of the restricted region: I would expect the coarse grid to carry zero reduction weight everywhere (as it does if the shiftout is one), but the boundaries have a non-zero weight instead, which leads to double counting, and hence the difference between domain volume and weight sum.

Attachments (1)

weight.par (4.4 KB) - added by Eloisa Bentivegna 6 years ago.

Download all attachments as: .zip

Change History (14)

Changed 6 years ago by Eloisa Bentivegna

Attachment: weight.par added

comment:1 Changed 6 years ago by Erik Schnetter

I believe this issue is unrelated to #1222. Looking at the weights, I see the effect you describe: coarse grid outer boundaries have non-zero weight where fine grid boundaries also have non-zero weight. This seems to be an error in setting up the boundary weights, which apparently doesn't take refined grids correctly into account.

comment:2 Changed 6 years ago by Erik Schnetter

After investigating, the reason for this problem seems to be that the regions do not "wrap around". That is, a fine grid abutting the right boundary does not influence the left boundary of the domain. To remedy this, one would need to teach the weight setup explicitly about periodic boundaries.

Can you switch to cell centred grids? Would this work for you? These should not have this problems.

comment:3 Changed 6 years ago by Erik Schnetter

My last comment is wrong since periodic boundaries are not involved.

comment:4 Changed 6 years ago by Eloisa Bentivegna

Yes, this problem occurs whether periodic boundaries are involved or not.

I'm not sure cell-centred grids would work for me (possibly yes). It would be nice if the issue with the vertex-centred ones was solved though. Can I help in any way?

comment:5 Changed 6 years ago by Erik Schnetter

I believe now that the issue is simply that fine grid boundaries not taken into account when setting the weight for coarse grid boundaries. Thus coarse grid boundaries always have non-zero weight. This is probably a left-over assumption from when the fine grid was never near a boundary.

If that is the case, then addressing it will be non-trivial.

comment:6 Changed 6 years ago by Eloisa Bentivegna

I see. An observation is that the behavior is completely correct when using a shiftout of one (in that case all the coarse grid points are masked out), so perhaps it's not so hard to find out where this parameter makes a difference and fix it. I briefly attempted doing this before opening the ticket, but with no success. I'll give it another try.

comment:7 Changed 6 years ago by Erik Schnetter

I'm afraid it is not simple. With a shiftout of 1, this issue does not arise. If my reading of the code is correct, then we need to introduce a completely new mechanism for this, starting in CarpetLib/src/dh.cc.

comment:8 Changed 6 years ago by Eloisa Bentivegna

OK. I'll consider switching to cell-centred grids, but I may return on this later.

comment:9 Changed 5 years ago by Erik Schnetter

Milestone: ET_2013_05

comment:10 Changed 5 years ago by Frank Löffler

Has this a chance of getting fixed soon?

comment:11 Changed 5 years ago by Erik Schnetter

Eloisa: Is this still an issue (if you use cell-centred grids)?

Frank: No, this requires a non-trivial calculation to be corrected. This involves a special case that only occurs for periodic boundaries (vertex centring with boundary_shiftout=0).

comment:12 Changed 5 years ago by Eloisa Bentivegna

I haven't tried cell-centred grids yet. In any case, I would leave this ticket open as a low-priority enhancement request, rather than closing it and forgetting about the fact that this setup is not supported.

comment:13 Changed 5 years ago by Frank Löffler

Milestone: ET_2013_05

Unlikely to be fixed so late and still get tested.

Modify Ticket

Change Properties
Set your email in Preferences
Action
as new The owner will remain Erik Schnetter.
Next status will be 'review'.
as The resolution will be set.
to The owner will be changed from Erik Schnetter to the specified user.
Next status will be 'confirmed'.
The owner will be changed from Erik Schnetter to anonymous.

Add Comment


E-mail address and name can be saved in the Preferences.

 
Note: See TracTickets for help on using tickets.