-
Notifications
You must be signed in to change notification settings - Fork 141
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Spin density is incorrect in the batched driver port #4533
Comments
I don't see any stat.h5 files in the zip. What does e.g. h5ls give as the contents of the file and of the SpinDensity entry? I am most curious to see if both spin up and spin down components are present in the file itself. |
I didn't include the
|
Thanks. Just as a final sanity check can you verify that the data values inside the file are identical for up and down for the batched case? (this removes any possibility of issues with the intervening script) |
OK, that's pretty convincing. Clearly there is an odd plumbing or evaluation issue in QMCPACK. |
I've expanded the the SpinDensityNew estimator unit tests and the problem does not appear to be with it. |
Are you able to reproduce the problem? Where could I look (or what tests can I run) to isolate the source of the issue? |
@PDoakORNL Perhaps the issue is with the internal accumulation or similar? @aannabe Perhaps put all the files (wavefunction, etc) in a shared space at OLCF? This way others can rerun simply to reproduce. |
The current spin density integration tests won't catch this one because those tests have identical up/down electron counts. |
No I have not reproduced it yet. The problem does not appear to be at the estimator level, I'll PR the unit test with the bugfix right now it just reveals that the estimator has no problem accumulating unequal numbers of up and down electrons. But there is a lot of plumbing between the actual particle movses and between there and the hdf5 write. I'm getting q-e working right now to produce some wavefunctions. |
Thanks. I transferred my files to OLCF in case they would be needed: Please let me know if there are problems with permissions etc. |
I have been taking a look recently at some of the estimators, so I decided to take a stab at this. I found the bug, which is a simple one line fix. I don't have any unit tests written up, but I can PR the fix unless @PDoakORNL already has a PR ready |
I do have a PR ready, I have new unit testing to cover this issue. |
Unfortunately I'm still debugging the integration test with hdfarchive which has different requirements for successfully reading than I would expect and segfaults. Easy to say this was caused by me getting away with inadequate testing. And I have no idea yet what is wrong with hdfarchive if actually anything. |
@aannabe how quickly do you need this resolved? I recall you saying you have some jobs to submit very soon that need the spin density working? If this is time critical, there is a simple one line fix to correct the bug that can used as a quick fix for now while @PDoakORNL works on changing the estimators and removing ObservableHelper from Estimators. The problem with the current code is that qmcpack/src/Estimators/SpinDensityNew.cpp Line 230 in 20877c4
This would allow @aannabe to run now while the estimators get cleaned up in #4540 |
@camelto2 Thanks, I will use this fix to run my calculations for now. However, in my opinion, it might be better to get this one-liner in the develop branch so that others don't encounter the same problem. We can then close this issue and create a new one around removing |
I agree. A one-line fix is always the right size. I would recommend someone just PR the one-liner; the rest can be prioritized against other needs. |
@camelto2 if you want to make a PR with that one liner I will approve. Test coverage here is poor so I've gotten a bit bogged down in the integration testing between the estimators and hdf5. None of that output was validated at all except through application tests. If I didn't have a second PR coming I think its important to note that the one line fix with no accompanying unit test to catch regression or prove its correctness is insufficient. |
completely agree. |
Describe the bug
The integrated charge from the spin density is not correct when using the batched driver.
To Reproduce
An isolated Boron atom (N_up = 2, N_down = 1) was simulated in a large box using QE. The spin density was accumulated using the legacy and batched drivers. In the legacy case, N_up and N_down are coming out correctly. In the batched case, N_up is also correct, but N_down is incorrect. In this case, N_down values are identical to N_up, so N_up is erroneously accumulated for both spin channels. See the attached zip for inputs and partial output files. Any ideas?
See the below summary and plots:
Legacy Up:
Legacy Down:
Batched Up:
Batched Down:
Expected behavior
Both spin channels should integrate to correct N_e.
System:
NERSC/Perlmutter. Latest develop version. See attached files for further info.
spin_density_issue.zip
The text was updated successfully, but these errors were encountered: