Segmentation fault in FAC MBPT calculation when more configurations involved

[mira42780534]

Dear all,

I encountered a segmentation fault while running an MBPT calculation with FAC through the Python interface.
The error message from Slurm is:

/var/spool/slurmd/job847232/slurm_script: line 13: 2298263 Segmentation fault (core dumped) python3 FAC.py

The segmentation fault does not occur when I reduce the number of configurations included in the MBPT calculation. This suggests that the issue may be related to the Hamiltonian matrix dimension becoming too large, possibly for a particular symmetry block. Are there any suggested strategies to avoid this issue, such as changing MBPT settings, or increasing the Hamiltonian matrix size limit?

---

My python input file is

from pfac.fac import *
import os

InitializeMPI(32)

z = 50
k = 36
a = ATOMICSYMBOL[z]
SetAtom(a)
SetUTA(0)

Closed("1s 2s 2p 3s 3p 3d")
Config('4s2 4p6', group='4s2 4p6')
Config('4s2 4p5 4d1', group='4s2 4p5 4d1')
Config('4s2 4p5 4f1', group='4s2 4p5 4f1')
Config('4s1 4p6 4d1', group='4s1 4p6 4d1')
Config('4s1 4p6 4f1', group='4s1 4p6 4f1')

Config('4s2 4p5 5*1', group='4s2 4p5 5*1')

Config('4s2 4p4 4d2', group='4s2 4p4 4d2')
Config('4s2 4p4 4d1 4f1', group='4s2 4p4 4d1 4f1')
Config('4s2 4p4 4f2', group='4s2 4p4 4f2')
Config('4s1 4p5 4d2', group='4s1 4p5 4d2')
Config('4s1 4p5 4d1 4f1', group='4s1 4p5 4d1 4f1')
Config('4s1 4p5 4f2', group='4s1 4p5 4f2')

Config('4s2 4p3 4d3', group='4s2 4p3 4d3')
Config('4s2 4p3 4d2 4f1', group='4s2 4p3 4d2 4f1')
Config('4s2 4p3 4d1 4f2', group='4s2 4p3 4d1 4f2')
Config('4s2 4p3 4f3', group='4s2 4p3 4f3')
Config('4s1 4p4 4d2 4f1', group='4s1 4p4 4d2 4f1')
Config('4s1 4p4 4d1 4f2', group='4s1 4p4 4d1 4f2')

Config('4s2 4p5', group='4s2 4p5')

conf_list = ['4s2 4p6', '4s2 4p5 4d1', '4s2 4p5 4f1', '4s1 4p6 4d1', '4s1 4p6 4f1', '4s2 4p5 5*1',
             '4s2 4p4 4d2', '4s2 4p4 4d1 4f1', '4s2 4p4 4f2', '4s1 4p5 4d2', '4s1 4p5 4d1 4f1', '4s1 4p5 4f2',
             '4s2 4p3 4d3', '4s2 4p3 4d2 4f1', '4s2 4p3 4d1 4f2', '4s2 4p3 4f3', '4s1 4p4 4d2 4f1', '4s1 4p4 4d1 4f2',
             '4s2 4p5']

ConfigEnergy(0)
OptimizeRadial(['4s2 4p6', '4s2 4p5 4d1', '4s2 4p5 4f1', '4s2 4p5'])
ConfigEnergy(1)

k = 3
n1 = [4,5,6,7,8,9,10,11,12,13,14]
n2 = [4,5,6,7,8,9,10,11,12,13,14]

Structure(1, 2)
StructureMBPT('b.en', ['corr.h', 'ham.h'], conf_list, k, n1, n2, len(conf_list)) # Causing segmentation fault
# StructureMBPT('b.en', ['corr.h', 'ham.h'], conf_list, k, n1, n2, 2) # Successive run

MemENTable('b.en')
PrintTable('b.en', 'a.en', 1)

BasisTable('basis.ba')

FinalizeMPI()

[mfgu]

this is likely too large configuration space. I doubt even a regular CI calculation can handle all those 4d3 and 4f3 configurations.

[mira42780534]

@mfgu Thank you for your reply.

I would like to provide two additional observations:

1. I checked the memory usage during the calculation, and it does not seem to be the limiting factor. The job only used about 30 GB of RAM, while the node has 2 TB of RAM available. Is there any way to increase the internal limit on the configuration space or Hamiltonian block size in FAC?

2. I can now run the input file after recompiling FAC. My machine has both gcc/11.2.0 and gcc/8.5.0 available. When FAC is compiled with gcc/11.2.0, the same input file leads to a segmentation fault. However, when FAC is compiled with gcc/8.5.0, the calculation can run successfully for this case.

That said, even with the gcc/8.5.0 build, I can still encounter a segmentation fault when the configuration space is made larger. Therefore, it seems that the crash may be related both to the compiler/build environment and to the size of the configuration/Hamiltonian space.

Do you have any suggestions on how to diagnose this further, or whether there are compile-time parameters or internal array-size limits that I should modify?

Thank you very much.

[mfgu]

I don't have such big machines to run your example. can you compile with debug mode, get the core dump, then load it in gdb, find out where the segfault is triggered.

[maj-mas]

I've also noticed these crashes for what I would consider modest sets of configurations. What is the suggested way to compile for a debug target?

[mfgu]

try to reduce the configuration list size, find the minimum set that triggers the segfault, maybe I can work from that.