Ascent crash: Floating point exception

MFIX-Exa
21

also, as you probably noticed, when it crashed it printed the first png file but not the second. which is strange, because the second image is just a subset of the first image. same slice, subset of particles thresholded by phase. it has to be something in that threshold, likely at the mass inflow.

22

Thanks @wfullmer . Your run was using 72 MPI procs, mine was using 128, so that possibly explains the difference in the images. May I ask what was the reason for choosing 72? This case has 512 boxes so I thought it would be good to use a number that divides 512.

I also noticed that the s2/p2 messages are somehow buffered and donā€™t appear until the end. Iā€™m hoping to fix that too, but thatā€™s a separate problem.

Iā€™m currently running some debug-mode jobs on Joule under gdb and hoping to catch a full stack trace with local variables, etc, as the Viskores team asked for. However these are running extremely slowly. What weā€™re trying to catch is the failing assert in Viskores:

ArrayPortalBasic.h:73:
Assertion `index < this->NumberOfValues' failed.

not the ā€œs2/p2ā€ message (which may or may not be related to the crash)

Will follow up when I get more info.

23

I just guessed that 6^3 grids was too small and used a larger size. Specifically, 8x8x16 because that fit on one node. This is why I asked if Yupeng set it up. As one of the Bernoulliā€™s said of Newton, I recognize the lion by his paw.

24

For the record, Will ran with

amr.max_grid_size_x =    8
amr.max_grid_size_y =    8
amr.max_grid_size_z =   16

so itā€™s not the same inputs file. Larger max_grid_size was one of my recommendations in a previous post. But Iā€™m trying to debug the reported Ascent/Viskores issue which is triggered by Femiā€™s original 6x6x6 settings (whether or not those are good settings, it should not cause a crash).

25

Update on the Ascent crash. Iā€™ve identified the problem and submitted a fix:

Threshold filter on point meshes crashes due to CleanGrid merging coincident points Ā· Issue #1686 Ā· Alpine-DAV/ascent Ā· GitHub

Ascent mishandles coincident or very-close (within merge threshold) particles. This would probably only happen in a PIC simulation.

This was interesting, I had a chance to learn a lot about the internals of Ascent and Viskores.

@oyedejifemi if youā€™re still trying to run this simulation, and donā€™t want to wait for Ascent to fix this, you can clone my repo at github.com/charlesgwaldman/ascent and check out the threshold-no-point-merge branch. Please let me know if this works for you. Iā€™ve run your job to tstop=10 without errors.

ā€“ Charles

1 Like
26