def run(custom_command=''):
"""
Main call of the function
It recovers the initialised instances of cosmo Class, :class:`Data` and the
NameSpace containing the command line arguments, feeding into the sampler.
Parameters
----------
custom_command: str
allows for testing the code
"""
# Create all the instances of the needed classes to run the code. The safe
# initialisation handles the errors.
cosmo, data, command_line, success = safe_initialisation(
custom_command)
# If success is False, it means either that the initialisation was not
# successful, or that it was simply an analysis call. The run should
# stop
if not success:
return
# Once that the initialisation phase is done, one can import the sampler
import sampler
# Generic sampler call
sampler.run(cosmo, data, command_line)
return
def run(custom_command=''):
"""
Main call of the function
It recovers the initialised instances of cosmo Class, :class:`Data` and the
NameSpace containing the command line arguments, feeding into the sampler.
Parameters
----------
custom_command: str
allows for testing the code
"""
# Create all the instances of the needed classes to run the code. The safe
# initialisation handles the errors.
cosmo, data, command_line, success = safe_initialisation(
custom_command)
# If success is False, it means either that the initialisation was not
# successful, or that it was simply an analysis call. The run should
# stop
if not success:
return
# Once that the initialisation phase is done, one can import the
# sampler
import sampler
# Generic sampler call
sampler.run(cosmo, data, command_line)
return
def run(custom_command=''):
"""
Main call of the function
It recovers the initialised instances of cosmo Class, :class:`Data` and the
NameSpace containing the command line arguments, feeding into the sampler.
.. note::
A possible parallelization would take place here.
Parameters
----------
custom_command: str
allows for testing the code
"""
# Initialisation routine
cosmo, data, command_line, success = initialise(custom_command)
# If success is False, it means either that the initialisation was not
# successful, or that it was simply an analysis call. The run should stop
if not success:
return
# Once that the initialisation phase is done, one can import the sampler
import sampler
# Generic sampler call
sampler.run(cosmo, data, command_line)
return
def mpi_run(custom_command=""):
"""
Launch a simple MPI run, with no communication of covariance matrix
It simply allows the first process to create the folder - so that the
log.param is properly written. A signal is then send to the other
processes, that contains the chain number of the parent run.
In order to be sure to have different chain numbers, it adds the rank of
the process and the initial job number - this should avoid conflict, but
can be subject of future improvements
"""
from mpi4py import MPI
comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()
# this will be the master cpu. This guy will create - or append - a folder,
# being sure to be the first to do so.
if rank == 0:
# First initialisation
cosmo, data, command_line, success = safe_initialisation(
custom_command, comm, nprocs)
regexp = re.match(".*__(\w*).txt", data.out_name)
suffix = regexp.groups()[0]
# Send an "OK" signal to all the other processes, actually giving the
# suffix of this master chain. All the other will add 1 to this number
for index in range(1, nprocs):
comm.send(suffix, dest=index, tag=1)
else:
# If the rank is not 0, it is a slave process. It waits to receive the
# "OK" message, which is immediatly discarded.
suffix = comm.recv(source=0, tag=1)
# If a failed message was passed, exit the process
if suffix == 'failed':
return
# Concatenate the rank to the suffix, and not the opposite, this should
# avoid any conflicting name
if not custom_command:
custom_command = " ".join(sys.argv[1:])
custom_command += " --chain-number %s" % str(rank)+suffix
cosmo, data, command_line, success = initialise(custom_command)
import sampler
sampler.run(cosmo, data, command_line)
return
def mpi_run(custom_command=""):
"""
Launch a simple MPI run, with no communication of covariance matrix
Each process will make sure to initialise the folder if needed. Then and
only then, it will send the signal to its next in line to proceed. This
allows for initialisation over an arbitrary cluster geometry (you can have
a single node with many cores, and all the chains living there, or many
nodes with few cores). The speed loss due to the time spend checking if the
folder is created should be negligible when running decently sized chains.
Each process will send the number that it found to be the first available
to its friends, so that the gathering of information post-run is made
easier. If a chain number is specified, this will be used as the first
number, and then incremented afterwards with the rank of the process.
"""
from mpi4py import MPI
comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()
success = True
folder = ''
# If the process is not the zeroth one, then wait for a signal from your
# n-1 before initializing the folder
if rank != 0:
status = comm.recv(source=rank-1, tag=1)
folder = comm.recv(source=rank-1, tag=2)
if status == 'failed':
success = False
else:
number = status
if success:
if not custom_command:
custom_command = " ".join(sys.argv[1:])
if rank != 0:
custom_command += " --chain-number %s" % str(int(number)+1)
# First check if the folder is there
already_sent = False
if rank != 0 and rank < nprocs-1:
status = int(number)+1
if Data.folder_is_initialised(folder):
comm.send(status, dest=rank+1, tag=1)
comm.send(folder, dest=rank+1, tag=2)
already_sent = True
# Then, properly initialise
cosmo, data, command_line, success = safe_initialisation(
custom_command, comm, nprocs)
# The first initialisation should check a few more things
if rank == 0:
# Check that the run asked is compatible with mpirun and prepare.
if command_line.subparser_name == 'info':
warnings.warn(
"Analyzing the chains is not supported in mpirun"
" so this will run on one core only.")
status = 'failed'
elif command_line.method == "MH":
regexp = re.match(".*__(\w*).txt", data.out_name)
suffix = regexp.groups()[0]
status = suffix
elif command_line.method == "NS":
status = 1
elif command_line.method == "PC":
status = 1
elif command_line.method == "NN":
status = 1
else:
warnings.warn(
"The method '%s' is not supported"%(command_line.method) +
" in mpirun so this will run on one core only.")
status = 'failed'
folder = data.out_name
elif rank < nprocs-1:
status = int(number)+1
# Send an "OK" signal to the next in line, giving the its own chain
# number for the other to add 1
if rank < nprocs-1:
if not already_sent:
comm.send(status, dest=rank+1, tag=1)
comm.send(folder, dest=rank+1, tag=2)
else:
if rank < nprocs-1:
comm.send('failed', dest=rank+1, tag=1)
comm.send('', dest=rank+1, tag=2)
if success:
import sampler
sampler.run(cosmo, data, command_line)
return
def mpi_run(custom_command=""):
"""
Launch a simple MPI run, with no communication of covariance matrix
Each process will make sure to initialise the folder if needed. Then and
only then, it will send the signal to its next in line to proceed. This
allows for initialisation over an arbitrary cluster geometry (you can have
a single node with many cores, and all the chains living there, or many
nodes with few cores). The speed loss due to the time spend checking if the
folder is created should be negligible when running decently sized chains.
Each process will send the number that it found to be the first available
to its friends, so that the gathering of information post-run is made
easier. If a chain number is specified, this will be used as the first
number, and then incremented afterwards with the rank of the process.
"""
from mpi4py import MPI
comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()
success = True
folder = ''
# If the process is not the zeroth one, then wait for a signal from your
# n-1 before initializing the folder
if rank != 0:
status = comm.recv(source=rank-1, tag=1)
folder = comm.recv(source=rank-1, tag=2)
if status == 'failed':
success = False
else:
number = status
if success:
if not custom_command:
custom_command = " ".join(sys.argv[1:])
if rank != 0:
custom_command += " --chain-number %s" % str(int(number)+1)
# First check if the folder is there
already_sent = False
if rank != 0 and rank < nprocs-1:
status = int(number)+1
if Data.folder_is_initialised(folder):
comm.send(status, dest=rank+1, tag=1)
comm.send(folder, dest=rank+1, tag=2)
already_sent = True
# Then, properly initialise
cosmo, data, command_line, success = safe_initialisation(
custom_command, comm, nprocs)
# The first initialisation should check a few more things
if rank == 0:
# Check that the run asked is compatible with mpirun and prepare.
if command_line.subparser_name == 'info':
warnings.warn(
"Analyzing the chains is not supported in mpirun"
" so this will run on one core only.")
status = 'failed'
elif command_line.method == "MH":
regexp = re.match(".*__(\w*).txt", data.out_name)
suffix = regexp.groups()[0]
status = suffix
elif command_line.method == "NS":
status = 1
else:
warnings.warn(
"The method '%s' is not supported"%(command_line.method) +
" in mpirun so this will run on one core only.")
status = 'failed'
folder = data.out_name
elif rank < nprocs-1:
status = int(number)+1
# Send an "OK" signal to the next in line, giving the its own chain
# number for the other to add 1
if rank < nprocs-1:
if not already_sent:
comm.send(status, dest=rank+1, tag=1)
comm.send(folder, dest=rank+1, tag=2)
else:
if rank < nprocs-1:
comm.send('failed', dest=rank+1, tag=1)
comm.send('', dest=rank+1, tag=2)
if success:
import sampler
sampler.run(cosmo, data, command_line)
return
def mpi_run(custom_command=""):
"""
Launch a simple MPI run, with no communication of covariance matrix
It simply allows the first process to create the folder - so that the
log.param is properly written. A signal is then send to the other
processes, that contains the chain number of the parent run.
In order to be sure to have different chain numbers, it adds the rank of
the process and the initial job number - this should avoid conflict, but
can be subject of future improvements
"""
from mpi4py import MPI
comm = MPI.COMM_WORLD
nprocs = comm.Get_size()
rank = comm.Get_rank()
# this will be the master cpu. This guy will create - or append - a folder,
# being sure to be the first to do so.
if rank == 0:
# First initialisation
cosmo, data, command_line, success = safe_initialisation(
custom_command, comm, nprocs)
# Check that the run asked is compatible with mpirun and prepare.
if command_line.subparser_name == 'info':
warnings.warn(
"Analyzing the chains is not supported in mpirun"
" so this will run on one core only.")
status = 'failed'
elif command_line.method == "MH":
regexp = re.match(".*__(\w*).txt", data.out_name)
suffix = regexp.groups()[0]
status = suffix
elif command_line.method == "NS":
status = 1
else:
warnings.warn(
"The method '%s' is not supported"%(command_line.method) +
" in mpirun so this will run on one core only.")
status = 'failed'
# Send an "OK" signal to all the other processes, actually giving the
# suffix of this master chain. All the other will add 1 to this number
for index in range(1, nprocs):
comm.send(status, dest=index, tag=1)
else:
# If the rank is not 0, it is a slave process. It waits to receive the
# "OK" message, which is immediatly discarded.
status = comm.recv(source=0, tag=1)
# If a failed message was passed, exit the process
if status == 'failed':
success = False
else:
# Concatenate the rank to the suffix, and not the opposite, this
# should avoid any conflicting name
if not custom_command:
custom_command = " ".join(sys.argv[1:])
suffix = status
custom_command += " --chain-number %s" % str(int(suffix)+rank)
cosmo, data, command_line, success = initialise(custom_command)
if success:
import sampler
sampler.run(cosmo, data, command_line)
success = comm.gather(success, root=0)
return
