SSHCustodian 0.2.6

This module depends on the Custodian class in the custodian project, which is a wrapper that manages jobs running on computing clusters. The custodian module is part of The Materials Project.

This module extends the Custodian class by creating the subclass SSHCustodian, which adds the functionality to copy the temporary directory created via monty to the scratch partitions on slave compute nodes, provided that the cluster’s file-system is configured in this way. The implementation invokes a sub-process to utilize the ssh executable installed on the cluster, so it is not particularly elegant or platform independent, nor is this solution likely to be general to all clusters. This is why this modification has not been submitted as a pull request to the main Custodian project.

You use SSHCustodian in the same way as the Custodian class, and it should integrate in with your existing scripts. The SSHCustodian class takes two additional arguments when creating a new instance, scratch_dir_node_only and pbs_nodefile:

scratch_dir_node_only (bool): If set to True, custodian will grab the list
    of nodes in the file path provided to pbs_nodefile and use copy the
    temp_dir to the scratch_dir on each node over ssh. This is necessary on
    cluster setups where each node has its own independent scratch
    partition.

pbs_nodefile (str): The filepath to the list of nodes to be used in a
    calculation. If this path does not point to a valid file, then
    scratch_dir_node_only will be automatically set to False.

The subclass SSHVaspJob was also created, which overrides the setup method to also check the environment variable PBS_NUM_PPN when auto_npar = True. This is necessary to implement for using compute node scratch partitions on PBS-based queueing systems, as the generic method of using multiprocessing.cpu_count() to count the number of cores will include hyperthreads, which will overestimate the number of physical cores and lead to NPAR being set too large. One consequence of setting NPAR too high is that a VASP job will hang, consuming resources but not doing anything useful. If you are using this kind of scratch directory, be careful about setting NPAR.

On many clusters, the filepath for the list of compute nodes is in the environment variable PBS_NODEFILE, which can be accessed in bash as $PBS_NODEFILE and in python using the os module. An example of how SSHCustodian can be used in a script is the following:

import logging
import os
from sshcustodian.sshcustodian import SSHCustodian
from custodian.vasp.handlers import (VaspErrorHandler,
                                     UnconvergedErrorHandler,
                                     MeshSymmetryErrorHandler,
                                     NonConvergingErrorHandler,
                                     PotimErrorHandler)
from custodian.vasp.validators import VasprunXMLValidator
from sshcustodian.vasp.sshjobs import SSHVaspJob
from pymatgen.io.vasp import VaspInput

FORMAT = '%(asctime)s %(message)s'
logging.basicConfig(format=FORMAT, level=logging.INFO, filename="run.log")

class VaspInputArgs:
    def __init__(self):
        """
        Set the default values for running a VASP job.
        """
        self.static_kpoint = 1

    def import_dict(self, in_dict):
        """
        Create and update self variables using dictionary.
        """
        for (key, value) in iteritems(in_dict):
            if key == "command":
                self.command = value
            if key == "static_kpoint":
                self.static_kpoint = value
            if key == "jobs":
                self.jobs = value


def get_runs(args):
    vasp_command = args.command.split()
    njobs = len(args.jobs)
    for i, job in enumerate(args.jobs):
        final = False if i != njobs - 1 else True
        if any(c.isdigit() for c in job):
            suffix = "." + job
        else:
            suffix = ".{}{}".format(job, i + 1)
        settings = []
        backup = True if i == 0 else False
        copy_magmom = False
        vinput = VaspInput.from_directory(".")
        if i > 0:
            settings.append(
                {"file": "CONTCAR",
                 "action": {"_file_copy": {"dest": "POSCAR"}}})
        job_type = job.lower()
        auto_npar = True
        if job_type.startswith("static"):
            m = [i * args.static_kpoint for i in vinput["KPOINTS"].kpts[0]]
            settings.extend([
                {"dict": "INCAR",
                 "action": {"_set": {"NSW": 0}}},
                {'dict': 'KPOINTS',
                 'action': {'_set': {'kpoints': [m]}}}])

        yield SSHVaspJob(vasp_command, final=final, suffix=suffix,
                         backup=backup, settings_override=settings,
                         copy_magmom=copy_magmom, auto_npar=auto_npar)


logging.info("Handlers used are %s" % args.handlers)
scratch_root = os.path.abspath("/scratch")
pbs_nodefile = os.environ.get("PBS_NODEFILE")
job_args = VaspInputArgs()
job_dict = {"command": "pvasp",
            "jobs": ["static"]}
job_args.import_dict(job_dict)
handlers = [VaspErrorHandler(), MeshSymmetryErrorHandler(),
            UnconvergedErrorHandler(), NonConvergingErrorHandler(),
            PotimErrorHandler()]
validators = [VasprunXMLValidator()]
c = SSHCustodian(handlers, get_runs(job_args), validators,
                 checkpoint=True,
                 scratch_dir=scratch_root,
                 scratch_dir_node_only=True,
                 pbs_nodefile=pbs_nodefile)
c.run()

Note that depending on how your cluster is configured, the "command": "pvasp" will need to be changed to however you invoke a parallel job.

For further information on how to use custodian, consult the custodian project documentation.