One benefit of the proposal method used in nessai
is it allows for simple parallelisation of the likelihood evaluation since new live points are drawn in batches and then stored. The likelihood can therefore be precomputed and stored later use.
Likelihood parallelisation can be enabled in nessai
by setting the keyword argument n_pool
when calling FlowSampler
. This determines the size of the multiprocessing pool to use for evaluating the likelihood.
Note
If running nessai
via a job scheduler such as HTCondor, remember to set the number of requested CPUs accordingly.
Alternatively, nessai
can use a user-defined pool. This is specified by setting the pool
argument in NestedSampler
or FlowSampler
. Some variables must be initialised when creating the pool, this is done using :py~nessai.utils.multiprocessing.initialise_pool_variables
:
from multiprocessing import Pool
from nessai.utils.multiprocessing import initialise_pool_variables
model = GaussianModel()
pool = Pool(
processes=2,
initializer=initialise_pool_variables,
initargs=(model,),
)
pool
can then passed to the pool
keyword argument when setting up the sampler.
ray
includes a distributed multiprocessing pool that can also be used with nessai
. Simply import ray.util.multiprocessing.Pool
instead of the standard pool and initialise using the method described above.
When a pool object is passed to nessai
it tries to determine how many processes the pool contains and (if the likelihood is vectorised) uses this information to determine the chunk size when evaluating the likelihood. If it can not determine this, then likelihood vectorisation will be disabled. This can be avoided by specifying n_pool
and max_threads
when initialising the sampler.
PyTorch supports different forms of parallelisation (see the PyTorch documentation for details). In nessai
, the user can configured the number of threads used for intra-op parallelisation by specifying the pytorch_threads
argument in FlowSampler
. This value does not have to match the number of threads use for the multiprocessing pool. By default, it is set to 1 to avoid all available resources being used.
Note
Scaling with pytorch_threads
can vary greatly between different systems and installations of PyTorch. We recommended testing different values before running large-scale analyses.
../examples/parallelisation_example.py
- :py
nessai.utils.threading.configure_threads
- :py
nessai.utils.multiprocessing.initialise_pool_variables
- :py
nessai.model.Model.configure_pool
- :py
nessai.model.Model.close_pool