def mpi_supervisor(hostfile, args):
"""Function used by the MPI supervisor to get the information needed to
form the mpiexec call to run an MPI version of the program
"""
import os
import stat
import sys
from metawards.utils import Console
Console.print("RUNNING AN MPI PROGRAM")
cores_per_node = get_cores_per_node(args)
Console.print(
f"Will run jobs assuming {cores_per_node} cores per compute node")
# based on the number of threads requested and the number of cores
# per node, we can work out the number of mpi processes to start,
# and can write a hostfile that will create the right layout
nthreads = get_threads_per_task(args)
Console.print(f"Will use {nthreads} OpenMP threads per model run...")
tasks_per_node = int(cores_per_node / nthreads)
Console.print(f"...meaning that the number of model runs per node will be "
f"{tasks_per_node}")
# Next, read the hostfile to get a unique list of hostnames
hostnames = {}
with open(hostfile, "r") as FILE:
line = FILE.readline()
while line:
hostname = line.strip()
if len(hostname) > 0:
hostnames[hostname] = 1
line = FILE.readline()
hostnames = list(hostnames.keys())
hostnames.sort()
Console.print(f"Number of compute nodes equals {len(hostnames)}")
Console.print(", ".join(hostnames))
# how many tasks can we perform in parallel?
nprocs = tasks_per_node * len(hostnames)
if args.nprocs:
if nprocs != args.nprocs:
Console.print(f"WARNING: You are using an unrecommended number of "
f"processes {args.nprocs} for the cluster {nprocs}.")
nprocs = args.nprocs
Console.print(
f"Total number of parallel processes to run will be {nprocs}")
Console.print(f"Total number of cores in use will be {nprocs*nthreads}")
# Now write a new hostfile that round-robins the MPI tasks over
# the nodes for 'tasks_per_node' runs
hostfile = f"_metawards_hostfile_{os.getpid()}"
Console.print(f"Writing hostfile to {hostfile}")
with open(hostfile, "w") as FILE:
i = 0
while i < nprocs:
for hostname in hostnames:
FILE.write(hostname + "\n")
i += 1
if i == nprocs:
break
# now craft the mpiexec command that will use this hostfile to
# run the job - remember to pass the option to stop the main process
# attempt to become a supervisor itself...
mpiexec = os.getenv("MPIEXEC")
if mpiexec is None:
mpiexec = "mpiexec"
# check for weird mpiexecs...
import subprocess
import shlex
try:
args = shlex.split(f"{mpiexec} -v")
p = subprocess.run(args,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT)
v = p.stdout.decode("utf-8").strip()
Console.print(f"{mpiexec} -v => {v}")
if v.find("HPE HMPT") != -1:
raise ValueError(
"metawards needs a more modern MPI library than HPE's, "
"so please compile to another MPI and use that.")
except Exception as e:
Console.error(f"[ERROR] {e.__class__} {e}")
pyexe = sys.executable
script = os.path.abspath(sys.argv[0])
args = " ".join(sys.argv[1:])
cmd = f"{mpiexec} -np {nprocs} -hostfile {hostfile} " \
f"{pyexe} -m mpi4py {script} --already-supervised {args} " \
f"--nprocs {nprocs}"
Console.print("Executing MPI job using")
Console.command(cmd)
try:
args = shlex.split(cmd)
subprocess.run(args).check_returncode()
except Exception as e:
Console.error("ERROR: Something went wrong!")
Console.error(f"{e.__class__}: {e}")
sys.exit(-1)
# clean up the hostfile afterwards... (we leave it if something
# went wrong as it may help debugging)
os.unlink(hostfile)
Console.print("MPI processes completed successfully")
def cli():
"""Main function for the command line interface. This does one of three
things:
1. If this is the main process, then it parses the arguments and
runs and manages the jobs
2. If this is a worker process, then it starts up and waits for work
3. If this is a supervisor process, then it query the job scheduling
system for information about the compute nodes to use, and will then
set up and run a manager (main) process that will use those
nodes to run the jobs
"""
from metawards.utils import Console
# get the parallel scheme now before we import any other modules
# so that it is clear if mpi4py or scoop (or another parallel module)
# has been imported via the required "-m module" syntax
parallel_scheme = get_parallel_scheme()
if parallel_scheme == "mpi4py":
from mpi4py import MPI
comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()
if rank != 0:
# this is a worker process, so should not do anything
# more until it is given work in the pool
Console.print(f"Starting worker process {rank+1} of {nprocs-1}...")
return
else:
Console.print("Starting main process...")
elif parallel_scheme == "scoop":
Console.print("STARTING SCOOP PROCESS")
import sys
args, parser = parse_args()
if not args.already_supervised:
hostfile = get_hostfile(args)
if hostfile:
# The user has asked to run a parallel job - this means that this
# process is the parallel supervisor
if args.mpi:
mpi_supervisor(hostfile, args)
return
elif args.scoop:
scoop_supervisor(hostfile, args)
return
# neither is preferred - if scoop is installed then use that
try:
import scoop # noqa - disable unused warning
have_scoop = True
except Exception:
have_scoop = False
if have_scoop:
scoop_supervisor(hostfile, args)
return
# do we have MPI?
try:
import mpi4py # noqa - disable unused warning
have_mpi4py = True
except Exception:
have_mpi4py = False
if have_mpi4py:
mpi_supervisor(hostfile, args)
return
# we don't have any other option, just keep going and
# use multiprocessing - in this case we don't need a
# supervisor and this is the main process
# This is now the code for the main process
# WE NEED ONE OF these listed options;
should_run = False
for arg in [
args.input, args.repeats, args.disease, args.additional,
args.model, args.iterator, args.extractor, args.demographics,
args.mixer, args.mover
]:
if arg is not None:
should_run = True
break
if not should_run:
parser.print_help(sys.stdout)
sys.exit(0)
if args.repeats is None:
args.repeats = [1]
# import the parameters here to speed up the display of help
from metawards import Parameters, Network, Population, print_version_string
# print the version information first, so that there is enough
# information to enable someone to reproduce this run
print_version_string()
Console.rule("Initialise")
if args.input:
# get the line numbers of the input file to read
if args.line is None or len(args.line) == 0:
linenums = None
Console.print(f"* Using parameters from all lines of {args.input}",
markdown=True)
else:
from metawards.utils import string_to_ints
linenums = string_to_ints(args.line)
if len(linenums) == 0:
Console.error(f"You cannot read no lines from {args.input}?")
sys.exit(-1)
elif len(linenums) == 1:
Console.print(
f"* Using parameters from line {linenums[0]} of "
f"{args.input}",
markdown=True)
else:
Console.print(
f"* Using parameters from lines {linenums} of "
f"{args.input}",
markdown=True)
from metawards import VariableSets, VariableSet
variables = VariableSets.read(filename=args.input,
line_numbers=linenums)
else:
from metawards import VariableSets, VariableSet
# create a VariableSets with one null VariableSet
variables = VariableSets()
variables.append(VariableSet())
nrepeats = args.repeats
if nrepeats is None or len(nrepeats) < 1:
nrepeats = [1]
if len(nrepeats) > 1 and len(variables) != len(nrepeats):
Console.error(f"The number of repeats {len(nrepeats)} must equal the "
f"number of adjustable variable lines {len(variables)}")
raise ValueError("Disagreement in the number of repeats and "
"adjustable variables")
# ensure that all repeats are >= 0
nrepeats = [0 if int(x) < 0 else int(x) for x in nrepeats]
if sum(nrepeats) == 0:
Console.error(f"The number of the number of repeats is 0. Are you "
f"sure that you don't want to run anything?")
raise ValueError("Cannot run nothing")
if len(nrepeats) == 1 and nrepeats[0] == 1:
Console.print("* Performing a single run of each set of parameters",
markdown=True)
elif len(nrepeats) == 1:
Console.print(
f"* Performing {nrepeats[0]} runs of each set of parameters",
markdown=True)
else:
Console.print(
f"* Performing {nrepeats} runs applied to the parameters",
markdown=True)
variables = variables.repeat(nrepeats)
# working out the number of processes and threads...
from metawards.utils import guess_num_threads_and_procs
(nthreads,
nprocs) = guess_num_threads_and_procs(njobs=len(variables),
nthreads=args.nthreads,
nprocs=args.nprocs,
parallel_scheme=parallel_scheme)
Console.print(
f"\n* Number of threads to use for each model run is {nthreads}",
markdown=True)
if nprocs > 1:
Console.print(
f"* Number of processes used to parallelise model "
f"runs is {nprocs}",
markdown=True)
Console.print(
f"* Parallelisation will be achieved using {parallel_scheme}",
markdown=True)
# sort out the random number seed
seed = args.seed
if seed is None:
import random
seed = random.randint(10000, 99999999)
if seed == 0:
# this is a special mode that a developer can use to force
# all jobs to use the same random number seed (15324) that
# is used for comparing outputs. This should NEVER be used
# for production code
Console.warning("Using special mode to fix all random number"
"seeds to 15324. DO NOT USE IN PRODUCTION!!!")
else:
Console.print(f"* Using random number seed {seed}", markdown=True)
# get the starting day and date
start_day = args.start_day
if start_day < 0:
raise ValueError(f"You cannot use a start day {start_day} that is "
f"less than zero!")
start_date = None
if args.start_date:
try:
from dateparser import parse
start_date = parse(args.start_date).date()
except Exception:
pass
if start_date is None:
from datetime import date
try:
start_date = date.fromisoformat(args.start_date)
except Exception as e:
raise ValueError(f"Cannot interpret a valid date from "
f"'{args.start_date}'. Error is "
f"{e.__class__} {e}")
if start_date is None:
from datetime import date
start_date = date.today()
Console.print(f"* Day zero is {start_date.strftime('%A %B %d %Y')}",
markdown=True)
if start_day != 0:
from datetime import timedelta
start_day_date = start_date + timedelta(days=start_day)
Console.print(f"Starting on day {start_day}, which is "
f"{start_day_date.strftime('%A %B %d %Y')}")
else:
start_day_date = start_date
# now find the MetaWardsData repository as this will be needed
# for the repeat command line too
(repository,
repository_version) = Parameters.get_repository(args.repository)
Console.print(f"* Using MetaWardsData at {repository}", markdown=True)
if repository_version["is_dirty"]:
Console.warning("This repository is dirty, meaning that the data"
"has not been committed to git. This may make "
"this calculation very difficult to reproduce")
# now work out the minimum command line needed to repeat this job
args.seed = seed
args.nprocs = nprocs
args.nthreads = nthreads
args.start_date = start_date.isoformat()
args.repository = repository
# also print the source of all inputs
import configargparse
Console.rule("Souce of inputs")
p = configargparse.get_argument_parser("main")
Console.print(p.format_values())
# print out the command used to repeat this job
repeat_cmd = "metawards"
for key, value in vars(args).items():
if value is not None:
k = key.replace("_", "-")
if isinstance(value, bool):
if value:
repeat_cmd += f" --{k}"
elif isinstance(value, list):
repeat_cmd += f" --{k}"
for val in value:
v = str(val)
if " " in v:
repeat_cmd += f" '{v}''"
else:
repeat_cmd += f" {v}"
else:
v = str(value)
if " " in v:
repeat_cmd += f" --{k} '{v}''"
else:
repeat_cmd += f" --{k} {v}"
Console.rule("Repeating this run")
Console.print("To repeat this job use the command;")
Console.command(repeat_cmd)
Console.print("Or alternatively use the config.yaml file that will be "
"written to the output directory and use the command;")
Console.command("metawards -c config.yaml")
# load all of the parameters
try:
params = Parameters.load(parameters=args.parameters)
except Exception as e:
Console.warning(
f"Unable to load parameter files. Make sure that you have "
f"cloned the MetaWardsData repository and have set the "
f"environment variable METAWARDSDATA to point to the "
f"local directory containing the repository, e.g. the "
f"default is $HOME/GitHub/MetaWardsData")
raise e
# should we profile the code? (default no as it prints a lot)
profiler = None
if args.no_profile:
profiler = None
elif args.profile:
from metawards.utils import Profiler
profiler = Profiler()
# load the disease and starting-point input files
Console.rule("Disease")
if args.disease:
params.set_disease(args.disease)
else:
params.set_disease("ncov")
Console.rule("Model data")
if args.model:
params.set_input_files(args.model)
else:
params.set_input_files("2011Data")
# load the user-defined custom parameters
Console.rule("Custom parameters and seeds")
if args.user_variables:
custom = VariableSet.read(args.user_variables)
Console.print(f"Adjusting variables to {custom}")
custom.adjust(params)
else:
Console.print("Not adjusting any parameters...")
# read the additional seeds
if args.additional is None or len(args.additional) == 0:
Console.print("Not using any additional seeds...")
else:
for additional in args.additional:
Console.print(f"Loading additional seeds from {additional}")
params.add_seeds(additional)
# what to do with the 0 state?
stage_0 = "R"
if args.disable_star:
Console.print("Disabling the * state. Stage 0 is the one and "
"only E state.")
stage_0 = "disable"
elif args.star_is_E:
Console.print("Setting the * state as an additional E state.")
stage_0 = "E"
else:
Console.print("Setting the * state as an additional R state.")
stage_0 = "R"
params.stage_0 = stage_0
# extra parameters that are set
params.UV = args.UV
# set these extra parameters to 0
params.static_play_at_home = 0
params.play_to_work = 0
params.work_to_play = 0
params.daily_imports = 0.0
Console.rule("Parameters")
Console.print(params, markdown=True)
# the size of the starting population
population = Population(initial=args.population,
date=start_day_date,
day=start_day)
Console.rule("Building the network")
network = Network.build(params=params,
population=population,
max_nodes=args.max_nodes,
max_links=args.max_links,
profiler=profiler)
if args.demographics:
from metawards import Demographics
Console.rule("Specialising into demographics")
demographics = Demographics.load(args.demographics)
Console.print(demographics)
network = network.specialise(demographics,
profiler=profiler,
nthreads=nthreads)
Console.rule("Preparing to run")
from metawards import OutputFiles
from metawards.utils import run_models
outdir = args.output
if outdir is None:
outdir = "output"
if args.force_overwrite_output:
prompt = None
else:
from metawards import input
def prompt(x):
return input(x, default="y")
auto_bzip = True
if args.auto_bzip:
auto_bzip = True
elif args.no_auto_bzip:
auto_bzip = False
if args.iterator:
iterator = args.iterator
else:
iterator = None
if args.extractor:
extractor = args.extractor
else:
extractor = None
if args.mixer:
mixer = args.mixer
else:
mixer = None
if args.mover:
mover = args.mover
else:
mover = None
with OutputFiles(outdir,
force_empty=args.force_overwrite_output,
auto_bzip=auto_bzip,
prompt=prompt) as output_dir:
# write the config file for this job to output/config.yaml
Console.rule("Running the model")
CONSOLE = output_dir.open("console.log")
Console.save(CONSOLE)
lines = []
max_keysize = None
for key, value in vars(args).items():
if max_keysize is None:
max_keysize = len(key)
elif len(key) > max_keysize:
max_keysize = len(key)
for key, value in vars(args).items():
if value is not None:
key = key.replace("_", "-")
spaces = " " * (max_keysize - len(key))
if isinstance(value, bool):
if value:
lines.append(f"{key}:{spaces} true")
else:
lines.append(f"{key}:{spaces} false")
elif isinstance(value, list):
s_value = [str(x) for x in value]
lines.append(f"{key}:{spaces} [ {', '.join(s_value)} ]")
else:
lines.append(f"{key}:{spaces} {value}")
CONFIG = output_dir.open("config.yaml", auto_bzip=False)
lines.sort(key=str.swapcase)
CONFIG.write("\n".join(lines))
CONFIG.write("\n")
CONFIG.flush()
CONFIG.close()
lines = None
result = run_models(network=network,
variables=variables,
population=population,
nprocs=nprocs,
nthreads=nthreads,
seed=seed,
nsteps=args.nsteps,
output_dir=output_dir,
iterator=iterator,
extractor=extractor,
mixer=mixer,
mover=mover,
profiler=profiler,
parallel_scheme=parallel_scheme)
if result is None or len(result) == 0:
Console.print("No output - end of run")
return 0
Console.rule("End of the run", style="finish")
Console.save(CONSOLE)
return 0
def scoop_supervisor(hostfile, args):
"""Function used by the scoop supervisor to get the information needed to
form the scoop call to run a scoop version of the program
"""
import os
import stat
import sys
from metawards.utils import Console
Console.print("RUNNING A SCOOP PROGRAM")
cores_per_node = get_cores_per_node(args)
Console.print(
f"Will run jobs assuming {cores_per_node} cores per compute node")
# based on the number of threads requested and the number of cores
# per node, we can work out the number of scoop processes to start,
# and can write a hostfile that will create the right layout
nthreads = get_threads_per_task(args)
Console.print(f"Will use {nthreads} OpenMP threads per model run...")
tasks_per_node = int(cores_per_node / nthreads)
Console.print(f"...meaning that the number of model runs per node will be "
f"{tasks_per_node}")
# Next, read the hostfile to get a unique list of hostnames
hostnames = {}
with open(hostfile, "r") as FILE:
line = FILE.readline()
while line:
hostname = line.strip()
if len(hostname) > 0:
hostnames[hostname] = 1
line = FILE.readline()
hostnames = list(hostnames.keys())
hostnames.sort()
Console.print(f"Number of compute nodes equals {len(hostnames)}")
Console.print(", ".join(hostnames))
# how many tasks can we perform in parallel?
nprocs = tasks_per_node * len(hostnames)
if args.nprocs:
if nprocs != args.nprocs:
Console.warning(
f"You are using a not-recommended number of "
f"processes {args.nprocs} for the cluster {nprocs}.")
nprocs = args.nprocs
Console.print(
f"Total number of parallel processes to run will be {nprocs}")
Console.print(f"Total number of cores in use will be {nprocs*nthreads}")
# Now write a new hostfile that round-robins the MPI tasks over
# the nodes for 'tasks_per_node' runs
hostfile = f"_metawards_hostfile_{os.getpid()}"
Console.print(f"Writing hostfile to {hostfile}")
with open(hostfile, "w") as FILE:
i = 0
while i < nprocs:
for hostname in hostnames:
FILE.write(hostname + "\n")
i += 1
if i == nprocs:
break
# now craft the scoop command that will use this hostfile to
# run the job - remember to pass the option to stop the main process
# attempting to become a supervisor itself...
import subprocess
import shlex
pyexe = sys.executable
script = os.path.abspath(sys.argv[0])
args = " ".join(sys.argv[1:])
# also need to tell the main program the number of processes
# as it can't work it out itself
cmd = f"{pyexe} -m scoop --hostfile {hostfile} -n {nprocs} " \
f"{script} --already-supervised {args} --nprocs {nprocs}"
Console.print("Executing scoop job using")
Console.command(cmd)
try:
args = shlex.split(cmd)
subprocess.run(args).check_returncode()
except Exception as e:
Console.error("ERROR: Something went wrong!")
Console.error(f"{e.__class__}: {e}")
sys.exit(-1)
# clean up the hostfile afterwards... (we leave it if something
# went wrong as it may help debugging)
os.unlink(hostfile)
Console.print("Scoop processes completed successfully")
