def test_standard(self):
try:
from mpi4py import MPI
if MPI.COMM_WORLD.Get_size() == 1:
# mpi4py available but NOT called via mpirun
MPI = None
except ImportError:
# mpi4py not even present! Single node by default:
MPI = None
self.assertAlmostEqual(comp_utils.get_MPI(), MPI)
def _unit_computation(self):
if self.verbose and self.mpi_rank == 0:
print("Rank {} at custom _unit_computation".format(self.mpi_rank))
resp_2d_list, dc_vec_list = self.data
req_cores = self._cores
MPI = get_MPI()
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
cores = 1
else:
rank = 0
cores = self._cores
if self.verbose:
print(
'Rank {} starting computing on {} cores (requested {} cores)'.
format(rank, cores, req_cores))
if cores > 1:
values = []
for loops_2d, curr_vdc in zip(resp_2d_list, dc_vec_list):
values += [
joblib.delayed(self._map_function)(x, [curr_vdc])
for x in loops_2d
]
results = joblib.Parallel(n_jobs=cores)(values)
# Finished reading the entire data set
print('Rank {} finished parallel computation'.format(rank))
else:
if self.verbose:
print("Rank {} computing serially ...".format(rank))
# List comprehension vs map vs for loop?
# https://stackoverflow.com/questions/1247486/python-list-comprehension-vs-map
results = []
for loops_2d, curr_vdc in zip(resp_2d_list, dc_vec_list):
results += [
self._map_function(vector, curr_vdc) for vector in loops_2d
]
self._results = results
def __init__(self, h5_main, cores=None, max_mem_mb=4 * 1024, verbose=False):
"""
Parameters
----------
h5_main : :class:`~pyUSID.io.usi_data.USIDataset`
The USID main HDF5 dataset over which the analysis will be performed.
cores : uint, optional
How many cores to use for the computation. Default: all available cores - 2 if operating outside MPI context
max_mem_mb : uint, optional
How much memory to use for the computation. Default 1024 Mb
verbose : bool, Optional, default = False
Whether or not to print debugging statements
"""
super(Process, self).__init__(h5_main, verbose=verbose)
MPI = get_MPI()
if MPI is not None:
# If we came here then, the user has intentionally asked for multi-node computation
comm = MPI.COMM_WORLD
self.mpi_comm = comm
self.mpi_rank = comm.Get_rank()
self.mpi_size = comm.Get_size()
if verbose:
print("Rank {} of {} on {} sees {} logical cores on the socket".format(comm.Get_rank(), comm.Get_size(),
MPI.Get_processor_name(),
cpu_count()))
# First, ensure that cores=logical cores in node. No point being economical / considerate
cores = psutil.cpu_count()
# It is sufficient if just one rank checks all this.
if self.mpi_rank == 0:
print('Working on {} ranks via MPI'.format(self.mpi_size))
# Ensure that the file is opened in the correct comm or something
if h5_main.file.driver != 'mpio':
raise TypeError('The HDF5 file should have been opened with driver="mpio". Current driver = "{}"'
''.format(h5_main.file.driver))
"""
# Not sure how to check for this correctly
messg = None
try:
if h5_main.file.comm != comm:
messg = 'The HDF5 file should have been opened with comm=MPI.COMM_WORLD. Currently comm={}'
''.format(h5_main.file.comm)
except AttributeError:
messg = 'The HDF5 file should have been opened with comm=MPI.COMM_WORLD'
if messg is not None:
raise TypeError(messg)
"""
else:
if verbose:
print('No mpi4py found or script was not called via mpixexec / mpirun. '
'Assuming single node computation')
self.mpi_comm = None
self.mpi_size = 1
self.mpi_rank = 0
# self.verbose = verbose and self.mpi_rank == 0
if MPI is not None:
MPI.COMM_WORLD.barrier()
# Not sure if we need a barrier here.
# Saving these as properties of the object:
self._cores = None
self.__ranks_on_socket = 1
self.__socket_master_rank = 0
self._max_pos_per_read = None
self._max_mem_mb = None
self._set_memory_and_cores(cores=cores, mem=max_mem_mb)
# Now have to be careful here since the below properties are a function of the MPI rank
self.__start_pos = None
self.__rank_end_pos = None
self.__end_pos = None
self.__pixels_in_batch = None
# Determining the max size of the data that can be put into memory
# all ranks go through this and they need to have this value any
self.duplicate_h5_groups = []
self.partial_h5_groups = []
self.process_name = None # Reset this in the extended classes
self.parms_dict = None