The Python API for partitioning a model over distributed and local hardware is described here.
Load balancing generates a
domain_decomposition given an
and a description of the hardware on which the model will run. Currently Arbor provides
one load balancer,
partition_load_balance(), and more will be added over time.
If the model is distributed with MPI, the partitioning algorithm for cells is
distributed with MPI communication. The returned
describes the cell groups on the local MPI rank.
partition_load_balance(recipe, context, hints)¶
The algorithm counts the number of each cell type in the global model, then partitions the cells of each type equally over the available nodes. If a GPU is available, and if the cell type can be run on the GPU, the cells on each node are put into one large group to maximise the amount of fine grained parallelism in the cell group. Otherwise, cells are grouped into small groups that fit in cache, and can be distributed over the available cores. Optionally, provide a dictionary of
partition_hints for certain cell kinds, by default this dictionary is empty.
The partitioning assumes that all cells of the same kind have equal computational cost, hence it may not produce a balanced partition for models with cells that have a large variance in computational costs.
Provide a hint on how the cell groups should be partitioned.
partition_hint(cpu_group_size, gpu_group_size, prefer_gpu)¶
By default returns a partition hint with
1, i.e., each cell is put in its own group,
max, i.e., all cells are put in one group, and
True, i.e., GPU usage is preferred.
The size of the cell group assigned to CPU. Must be positive, else set to default value.
The size of the cell group assigned to GPU. Must be positive, else set to default value.
Whether GPU usage is preferred.
Get the maximum size of cell groups.
An example of a partition load balance with hints reads as follows:
import arbor # Get a communication context (with 4 threads, no GPU) context = arbor.context(threads=4, gpu_id=None) # Initialise a recipe of user defined type my_recipe with 100 cells. n_cells = 100 recipe = my_recipe(n_cells) # The hints perfer the multicore backend, so the decomposition is expected # to never have cell groups on the GPU, regardless of whether a GPU is # available or not. cable_hint = arb.partition_hint() cable_hint.prefer_gpu = False cable_hint.cpu_group_size = 3 spike_hint = arb.partition_hint() spike_hint.prefer_gpu = False spike_hint.cpu_group_size = 4 hints = dict([(arb.cell_kind.cable, cable_hint), (arb.cell_kind.spike_source, spike_hint)]) decomp = arb.partition_load_balance(recipe, context, hints)
As defined in Domain Decomposition a domain decomposition is a description of the distribution of the model over the available computational resources. Therefore, the following data structures are used to describe domain decompositions.
Enumeration used to indicate which hardware backend to execute a cell group on.
Use multicore backend.
Use GPU backend.
Setting the GPU back end is only meaningful if the cell group type supports the GPU backend.
Describes a domain decomposition and is soley responsible for describing the distribution of cells across cell groups and domains. It holds cell group descriptions (
groups) for cells assigned to the local domain, and a helper function (
gid_domain()) used to look up which domain a cell has been assigned to. The
domain_decompositionobject also has meta-data about the number of cells in the global model, and the number of domains over which the model is destributed.
The domain decomposition represents a division of all of the cells in the model into non-overlapping sets, with one set of cells assigned to each domain.
A function for querying the domain id that a cell is assigned to (using global identifier
The number of domains that the model is distributed over.
The index of the local domain. Always 0 for non-distributed models, and corresponds to the MPI rank for distributed runs.
The total number of cells in the local domain.
The total number of cells in the global model (sum of
num_local_cellsover all domains).