Segmentation fault on ANM calculations on large systems, despite of tones of avaliable RAM

[rmendez]

Dear all,

I am trying to perform ANM calculations on a tetrameric membrane protein (PDB ID 3J8H, chains A,C,E,G) using only CA atoms, which amount a total of 14640 atoms. In our unversity we have a computer cluster with up to 1 TB of RAM, so for testing purposes I am booking up to 800 Gb, but the jobs crashes producing a segmentation fault (and core dump). Here it is the script:

---

from prody import * import matplotlib as mpl mpl.use('Agg') import matplotlib.pylab as plab import matplotlib.pyplot as plt import numpy as np

RyR_closed = parsePDB('3J8H_ACEG.pdb') RyR_closed = RyR_closed.select('protein and name CA')

compute ANM modes

anm = ANM('RyR closed ANM analysis') anm.buildHessian(RyR_closed) anm.calcModes()

write Modes to be represented by VMD

to be read by parseNMD subroutine

writeNMD('RyR_closed.ANM.nmd',anm,RyR_closed) quit()

---

And here is the error:

@> 107796 atoms and 1 coordinate set(s) were parsed in 1.23s. @> Hessian was built in 65.85s. /netscr/EX/lsf/killdevil/lsbatch/1420483745.214510: line 8: 12861 Segmentation fault (core dumped) ./prody_anm.CA.py

However I know that the maximum memory used is about 29 Gb:

Exited with exit code 139.

Resource usage summary:

CPU time :                                   5526.03 sec.
Max Memory :                                 29 GB
Average Memory :                             28.32 GB
Total Requested Memory :                     800.00 GB
Delta Memory :                               771.00 GB
Max Swap :                                   31 GB
Max Processes :                              3
Max Threads :                                12
Run time :                                   4616 sec.
Turnaround time :                            4683 sec.

I see two possible problems (not being a python programer):

1) Python memory allocation for arrays is predetermined to a size below the one needed in the diagonalization process (Hessian computation is made extremely fast)

2) Or perhaps it is known that LAPACK/BPACK subroutines used as diagonalizer (the ones in C) are not able to handle this array size.

Can anyone help me to perform ANM on large systems ?

Thank you so much,

Raul

[cihankayacihan]

Lapack is dealing with those over scipy. I guess it will be due to diagonalizer. Currently, we have an effort to build gpu based eigenvalue decomposition for large molecules with magma. With 1.9, it will be online.

[jamesmkrieger]

This is still in progress. There's a GPU implementation pretty much complete but not integrated yet.

[jamesmkrieger]

You could maybe try using Dask. This package handles much larger arrays and splits the computations into parallel units. They have a singular value decomposition (https://docs.dask.org/en/latest/array-api.html?highlight=linalg#dask.array.linalg.svd) that you could use for the matrix decomposition maybe. I have also thought about incorporating it, but never got a chance.

Alternatively, you could try coarse-graining further. One option could be to take ever second or third residue or even larger gaps (see https://pubmed.ncbi.nlm.nih.gov/11913377/ where this seems to work up to 1/40 of residues using a shifting scheme) or take the average position of every few residues. Another option is more rigorous Markov methods for hierarchical coarse-graining (https://pubmed.ncbi.nlm.nih.gov/17691893/) although this might be more challenging.

Another alternative could be to use our new interface CryoDy to obtain a coarse grained elastic network directly from a cryo-EM map (see http://prody.csb.pitt.edu/tutorials/cryoem_tutorial/).

Best wishes James

[jamesmkrieger]

I was just trying this yesterday. I think the best option is to use buildHessian with sparse=True and calcModes with turbo=False. It may also help to limit the number of cpus.

What I did was the following:

In [1]: from prody import *

In [2]: two_cct = parsePDB("4v8r")
@> Connecting wwPDB FTP server RCSB PDB (USA).
@> pdb4v8r.ent.gz download failed. 4v8r does not exist on ftp.wwpdb.org.
@> PDB download via FTP completed (0 downloaded, 1 failed).
@> Downloading PDB files via FTP failed, trying HTTP.
@> WARNING 4v8r download failed ('https://files.rcsb.org/download/4V8R.pdb.gz' could not be opened for reading, invalid URL or no internet connection).
@> PDB download via HTTP completed (0 downloaded, 1 failed).
@> WARNING Trying to parse mmCIF file instead
@> WARNING Could not find _pdbx_branch_scheme in lines.
@> WARNING Could not find _pdbx_struct_mod_residue in lines.
@> WARNING Could not find _atom_site_anisotrop in lines.
@> WARNING No anisotropic B factors found
@> 128780 atoms and 1 coordinate set(s) were parsed in 2.64s.
@> Secondary structures were assigned to 10973 residues.

In [3]: anm = ANM("all calphas")

In [4]: from threadpoolctl import threadpool_limits

In [5]: with threadpool_limits(limits=6, user_api="blas"):
   ...:     anm.buildHessian(two_cct.ca, cutoff=20, gamma=1, sparse=True, kdtree=False)
   ...:     anm.calcModes(5, zeros=False, turbo=False)

I've found that limiting the number of threads can also speed up the calculations sometimes