def decompose_survey_data(first, second, attrs, logger, smoothing, domain_factor=2,
angular=False, return_cartesian=False):
"""
Perform a domain decomposition on survey data, returning the
domain-demposed position and weight arrays for each object in the
correlating pair.
The domain decomposition is based on the Cartesian coordinates of
the input data (assumed to be in sky coordinates).
Load balancing is required since the distribution in Cartesian space
will likely not be uniform.
The implementation follows:
1. Decompose the first source and balance the particle load, such that
the first source is evenly distributed across all ranks and the
objects are spatially tight on a given rank.
2. Decompose the second source, ensuring a given rank holds all
particles within the desired maximum separation.
Parameters
----------
first : CatalogSource
the first source we are correlating
second : CatalogSource
the second source we are correlating
attrs : dict
dict of parameters from the pair counting algorithm
logger :
the current active logger
smoothing :
the maximum Cartesian separation implied by the user's binning
domain_factor : int, optional
the factor by which we over-sample the mesh with cells in a given
direction; higher values can lead to better performance
angular : bool, optional
if ``True``, the Cartesian positions used in the domain
decomposition are on the unit sphere
return_cartesian : bool, optional
whether to return the pos as (ra, dec, z), or the Cartesian (x, y, z)
Returns
-------
(pos1, w1), (pos2, w2) : array_like
the (decomposed) set of positions and weights to correlate
"""
from nbodykit.transform import StackColumns
comm = first.comm
# either (ra,dec) or (ra,dec,redshift)
poscols = [attrs['ra'], attrs['dec']]
if not angular: poscols += [attrs['redshift']]
# determine processor division for domain decomposition
np = split_size_3d(comm.size)
if comm.rank == 0:
logger.info("using cpu grid decomposition: %s" %str(np))
# stack position and compute
pos1 = StackColumns(*[first[col] for col in poscols])
pos1, w1 = first.compute(pos1, first[attrs['weight']])
N1 = comm.allreduce(len(pos1))
# only need cosmo if not angular
cosmo = attrs.get('cosmo', None) if not angular else None
if not angular and cosmo is None:
raise ValueError("need a cosmology to decompose non-angular survey data")
cpos1, cpos1_min, cpos1_max, rdist1 = get_cartesian(comm, pos1, cosmo=cosmo)
# pass in comoving dist to Corrfunc instead of redshift
if not angular:
pos1[:,2] = rdist1
# set up position for second too
if second is not None:
# stack position and compute for "second"
pos2 = StackColumns(*[second[col] for col in poscols])
pos2, w2 = second.compute(pos2, second[attrs['weight']])
N2 = comm.allreduce(len(pos2))
# get comoving dist and boxsize
cpos2, cpos2_min, cpos2_max, rdist2 = get_cartesian(comm, pos2, cosmo=cosmo)
# pass in comoving distance instead of redshift
if not angular:
pos2[:,2] = rdist2
else:
pos2 = pos1
w2 = w1
N2 = N1
cpos2_min = cpos1_min
cpos2_max = cpos1_max
cpos2 = cpos1
# determine global boxsize
if second is None:
cpos_min = cpos1_min
cpos_max = cpos1_max
else:
cpos_min = numpy.min(numpy.vstack([cpos1_min, cpos2_min]), axis=0)
cpos_max = numpy.max(numpy.vstack([cpos1_max, cpos2_max]), axis=0)
boxsize = cpos_max - cpos_min
if comm.rank == 0:
logger.info("position variable range on rank 0 (max, min) = %s, %s" % (cpos_max, cpos_min))
# initialize the domain
# NOTE: over-decompose by factor of 2 to trigger load balancing
grid = [
numpy.linspace(cpos_min[0], cpos_max[0], domain_factor*np[0] + 1, endpoint=True),
numpy.linspace(cpos_min[1], cpos_max[1], domain_factor*np[1] + 1, endpoint=True),
numpy.linspace(cpos_min[2], cpos_max[2], domain_factor*np[2] + 1, endpoint=True),
]
domain = GridND(grid, comm=comm, periodic=False)
# balance the load
domain.loadbalance(domain.load(cpos1))
if comm.rank == 0:
logger.info("Load balance done")
# if we want to return cartesian, redefine pos
if return_cartesian:
pos1 = cpos1
pos2 = cpos2
# decompose based on cartesian positions
layout = domain.decompose(cpos1, smoothing=0)
pos1 = layout.exchange(pos1)
w1 = layout.exchange(w1)
# get the position/weight of the secondaries
if smoothing > boxsize.max() * 0.25:
pos2 = numpy.concatenate(comm.allgather(pos2), axis=0)
w2 = numpy.concatenate(comm.allgather(w2), axis=0)
else:
layout = domain.decompose(cpos2, smoothing=smoothing)
pos2 = layout.exchange(pos2)
w2 = layout.exchange(w2)
# log the decomposition breakdown
log_decomposition(comm, logger, N1, N2, pos1, pos2)
return (pos1, w1), (pos2, w2)
def decompose_survey_data(first, second, attrs, logger, smoothing, domain_factor=2,
angular=False, return_cartesian=False):
"""
Perform a domain decomposition on survey data, returning the
domain-demposed position and weight arrays for each object in the
correlating pair.
The domain decomposition is based on the Cartesian coordinates of
the input data (assumed to be in sky coordinates).
Load balancing is required since the distribution in Cartesian space
will likely not be uniform.
The implementation follows:
1. Decompose the first source and balance the particle load, such that
the first source is evenly distributed across all ranks and the
objects are spatially tight on a given rank.
2. Decompose the second source, ensuring a given rank holds all
particles within the desired maximum separation.
Parameters
----------
first : CatalogSource
the first source we are correlating
second : CatalogSource
the second source we are correlating
attrs : dict
dict of parameters from the pair counting algorithm
logger :
the current active logger
smoothing :
the maximum Cartesian separation implied by the user's binning
domain_factor : int, optional
the factor by which we over-sample the mesh with cells in a given
direction; higher values can lead to better performance
angular : bool, optional
if ``True``, the Cartesian positions used in the domain
decomposition are on the unit sphere
return_cartesian : bool, optional
whether to return the pos as (ra, dec, z), or the Cartesian (x, y, z)
Returns
-------
(pos1, w1), (pos2, w2) : array_like
the (decomposed) set of positions and weights to correlate
"""
from nbodykit.transform import StackColumns
comm = first.comm
# either (ra,dec) or (ra,dec,redshift)
poscols = [attrs['ra'], attrs['dec']]
if not angular: poscols += [attrs['redshift']]
# determine processor division for domain decomposition
np = split_size_3d(comm.size)
if comm.rank == 0:
logger.info("using cpu grid decomposition: %s" %str(np))
# stack position and compute
pos1 = StackColumns(*[first[col] for col in poscols])
pos1, w1 = first.compute(pos1, first[attrs['weight']])
N1 = comm.allreduce(len(pos1))
# only need cosmo if not angular
cosmo = attrs.get('cosmo', None) if not angular else None
if not angular and cosmo is None:
raise ValueError("need a cosmology to decompose non-angular survey data")
cpos1, cpos1_min, cpos1_max, rdist1 = get_cartesian(comm, pos1, cosmo=cosmo)
# pass in comoving dist to Corrfunc instead of redshift
if not angular:
pos1 = pos1.copy() # we need to overwrite it; dask doesn't always return a copy after 0.18.1
pos1[:,2] = rdist1
# set up position for second too
if second is not None:
# stack position and compute for "second"
pos2 = StackColumns(*[second[col] for col in poscols])
pos2, w2 = second.compute(pos2, second[attrs['weight']])
N2 = comm.allreduce(len(pos2))
# get comoving dist and boxsize
cpos2, cpos2_min, cpos2_max, rdist2 = get_cartesian(comm, pos2, cosmo=cosmo)
# pass in comoving distance instead of redshift
if not angular:
pos2 = pos2.copy() # we need to overwrite it; dask doesn't always return a copy after 0.18.1
pos2[:,2] = rdist2
else:
pos2 = pos1
w2 = w1
N2 = N1
cpos2_min = cpos1_min
cpos2_max = cpos1_max
cpos2 = cpos1
# determine global boxsize
if second is None:
cpos_min = cpos1_min
cpos_max = cpos1_max
else:
cpos_min = numpy.min(numpy.vstack([cpos1_min, cpos2_min]), axis=0)
cpos_max = numpy.max(numpy.vstack([cpos1_max, cpos2_max]), axis=0)
boxsize = cpos_max - cpos_min
if comm.rank == 0:
logger.info("position variable range on rank 0 (max, min) = %s, %s" % (cpos_max, cpos_min))
# initialize the domain
# NOTE: over-decompose by factor of 2 to trigger load balancing
grid = [
numpy.linspace(cpos_min[0], cpos_max[0], domain_factor*np[0] + 1, endpoint=True),
numpy.linspace(cpos_min[1], cpos_max[1], domain_factor*np[1] + 1, endpoint=True),
numpy.linspace(cpos_min[2], cpos_max[2], domain_factor*np[2] + 1, endpoint=True),
]
domain = GridND(grid, comm=comm, periodic=False)
# balance the load
domain.loadbalance(domain.load(cpos1))
if comm.rank == 0:
logger.info("Load balance done")
# if we want to return cartesian, redefine pos
if return_cartesian:
pos1 = cpos1
pos2 = cpos2
# decompose based on cartesian positions
layout = domain.decompose(cpos1, smoothing=0)
pos1 = layout.exchange(pos1)
w1 = layout.exchange(w1)
# get the position/weight of the secondaries
if smoothing > boxsize.max() * 0.25:
pos2 = numpy.concatenate(comm.allgather(pos2), axis=0)
w2 = numpy.concatenate(comm.allgather(w2), axis=0)
else:
layout = domain.decompose(cpos2, smoothing=smoothing)
pos2 = layout.exchange(pos2)
w2 = layout.exchange(w2)
# log the decomposition breakdown
log_decomposition(comm, logger, N1, N2, pos1, pos2)
return (pos1, w1), (pos2, w2)
def fof(source, linking_length, comm, periodic, domain_factor, logger):
"""
Run Friends-of-friends halo finder.
Friends-of-friends was first used by Davis et al 1985 to define
halos in hierachical structure formation of cosmological simulations.
The algorithm is also known as DBSCAN in computer science.
The subroutine here implements a parallel version of the FOF.
The underlying local FOF algorithm is from `kdcount.cluster`,
which is an adaptation of the implementation in Volker Springel's
Gadget and Martin White's PM. It could have been done faster.
Parameters
----------
source: CatalogSource
the input source of particles; must support 'Position' column;
``source.attrs['BoxSize']`` is also used
linking_length: float
linking length in data units. (Usually Mpc/h).
comm: MPI.Comm
The mpi communicator.
Returns
-------
minid: array_like
A unique label of each position. The label is not ranged from 0.
"""
from pmesh.domain import GridND
np = split_size_3d(comm.size)
nd = np * domain_factor
if periodic:
BoxSize = source.attrs.get('BoxSize', None)
if BoxSize is None:
raise ValueError(
"cannot compute FOF clustering of source without 'BoxSize' in ``attrs`` dict"
)
if numpy.isscalar(BoxSize):
BoxSize = [BoxSize, BoxSize, BoxSize]
left = [0, 0, 0]
right = BoxSize
else:
BoxSize = None
left = numpy.min(comm.allgather(
source['Position'].min(axis=0).compute()),
axis=0)
right = numpy.max(comm.allgather(
source['Position'].max(axis=0).compute()),
axis=0)
grid = [
numpy.linspace(left[0], right[0], nd[0] + 1, endpoint=True),
numpy.linspace(left[1], right[1], nd[1] + 1, endpoint=True),
numpy.linspace(left[2], right[2], nd[2] + 1, endpoint=True),
]
domain = GridND(grid, comm=comm, periodic=periodic)
Position = source.compute(source['Position'])
np = comm.allgather(len(Position))
if comm.rank == 0:
logger.info(
"Number of particles max/min = %d / %d before spatial decomposition"
% (max(np), min(np)))
# balance the load
domain.loadbalance(domain.load(Position))
layout = domain.decompose(Position, smoothing=linking_length * 1)
np = comm.allgather(layout.recvlength)
if comm.rank == 0:
logger.info(
"Number of particles max/min = %d / %d after spatial decomposition"
% (max(np), min(np)))
comm.barrier()
minid = _fof_local(layout, Position, BoxSize, linking_length, comm)
comm.barrier()
minid = _fof_merge(layout, minid, comm)
return minid
def fof(source, linking_length, comm, periodic, domain_factor, logger):
"""
Run Friends-of-friends halo finder.
Friends-of-friends was first used by Davis et al 1985 to define
halos in hierachical structure formation of cosmological simulations.
The algorithm is also known as DBSCAN in computer science.
The subroutine here implements a parallel version of the FOF.
The underlying local FOF algorithm is from `kdcount.cluster`,
which is an adaptation of the implementation in Volker Springel's
Gadget and Martin White's PM. It could have been done faster.
Parameters
----------
source: CatalogSource
the input source of particles; must support 'Position' column;
``source.attrs['BoxSize']`` is also used
linking_length: float
linking length in data units. (Usually Mpc/h).
comm: MPI.Comm
The mpi communicator.
Returns
-------
minid: array_like
A unique label of each position. The label is not ranged from 0.
"""
from pmesh.domain import GridND
np = split_size_3d(comm.size)
nd = np * domain_factor
if periodic:
BoxSize = source.attrs.get('BoxSize', None)
if BoxSize is None:
raise ValueError("cannot compute FOF clustering of source without 'BoxSize' in ``attrs`` dict")
if numpy.isscalar(BoxSize):
BoxSize = [BoxSize, BoxSize, BoxSize]
left = [0, 0, 0]
right = BoxSize
else:
BoxSize = None
left = numpy.min(comm.allgather(source['Position'].min(axis=0).compute()), axis=0)
right = numpy.max(comm.allgather(source['Position'].max(axis=0).compute()), axis=0)
grid = [
numpy.linspace(left[0], right[0], nd[0] + 1, endpoint=True),
numpy.linspace(left[1], right[1], nd[1] + 1, endpoint=True),
numpy.linspace(left[2], right[2], nd[2] + 1, endpoint=True),
]
domain = GridND(grid, comm=comm, periodic=periodic)
Position = source.compute(source['Position'])
np = comm.allgather(len(Position))
if comm.rank == 0:
logger.info("Number of particles max/min = %d / %d before spatial decomposition" % (max(np), min(np)))
# balance the load
domain.loadbalance(domain.load(Position))
layout = domain.decompose(Position, smoothing=linking_length * 1)
np = comm.allgather(layout.newlength)
if comm.rank == 0:
logger.info("Number of particles max/min = %d / %d after spatial decomposition" % (max(np), min(np)))
comm.barrier()
minid = _fof_local(layout, Position, BoxSize, linking_length, comm)
comm.barrier()
minid = _fof_merge(layout, minid, comm)
return minid
