Hardware Management

Arbor provides two ways for working with hardware resources:

  • Prescribe the hardware resources and their contexts for use in Arbor simulations.
  • Query available hardware resources (e.g. the number of available GPUs), and initializing MPI.

Available Resources

Helper functions for checking cmake or environment variables, as well as configuring and checking MPI are the following:

arbor.config()

Returns a dictionary to check which options the Arbor library was configured with at compile time:

  • ARB_MPI_ENABLED
  • ARB_WITH_MPI4PY
  • ARB_GPU_ENABLED
  • ARB_VERSION
import arbor
arbor.config()

{'mpi': True, 'mpi4py': True, 'gpu': False, 'version': '0.2.1-dev'}
arbor.mpi_init()

Initialize MPI with MPI_THREAD_SINGLE, as required by Arbor.

arbor.mpi_is_initialized()

Check if MPI is initialized.

class arbor.mpi_comm
mpi_comm()

By default sets MPI_COMM_WORLD as communicator.

mpi_comm(object)

Converts a Python object to an MPI Communicator.

arbor.mpi_finalize()

Finalize MPI by calling MPI_Finalize.

arbor.mpi_is_finalized()

Check if MPI is finalized.

Prescribed Resources

The Python wrapper provides an API for:

  • prescribing which hardware resources are to be used by a simulation using proc_allocation.
  • opaque handles to hardware resources used by simulations called context.
class arbor.proc_allocation

Enumerates the computational resources on a node to be used for a simulation, specifically the number of threads and identifier of a GPU if available.

proc_allocation()

By default selects one thread and no GPU.

proc_allocation(threads, gpu_id)

Constructor that sets the number of threads and the id gpu_id of the available GPU.

threads

The number of CPU threads available, 1 by default.

gpu_id

The identifier of the GPU to use. Must be None, or a non-negative integer.

The gpu_id corresponds to the int device parameter used by CUDA API calls to identify gpu devices. Set to None to indicate that no GPU device is to be used. See cudaSetDevice and cudaDeviceGetAttribute provided by the CUDA API.

has_gpu()

Indicates whether a GPU is selected (i.e., whether gpu_id is None).

Here are some examples of how to create a proc_allocation.

import arbor

# default: one thread and no GPU selected
alloc1 = arbor.proc_allocation()

# 8 threads and no GPU
alloc2 = arbor.proc_allocation(8, None)

# reduce alloc2 to 4 threads and use the first available GPU
alloc2.threads = 4
alloc2.gpu_id  = 0
class arbor.context

An opaque handle for the hardware resources used in a simulation. A context contains a thread pool, and optionally the GPU state and MPI communicator. Users of the library do not directly use the functionality provided by context, instead they configure contexts, which are passed to Arbor interfaces for domain decomposition and simulation.

context()

Construct a local context with one thread, no GPU, no MPI.

context(alloc)

Create a local context, with no distributed/MPI, that uses the local resources described by proc_allocation.

alloc

The computational resources, one thread and no GPU by default.

context(alloc, mpi)

Create a distributed context, that uses the local resources described by proc_allocation, and uses the MPI communicator for distributed calculation.

alloc

The computational resources, one thread and no GPU by default.

mpi

The MPI communicator (see mpi_comm). mpi must be None, or an MPI communicator.

context(threads, gpu_id)

Create a context that uses a set number of threads and the GPU with id gpu_id.

threads

The number of threads available locally for execution, 1 by default.

gpu_id

The identifier of the GPU to use, None by default. Must be None, or a non-negative integer.

context(threads, gpu_id, mpi)

Create a context that uses a set number of threads and gpu identifier gpu_id and MPI communicator mpi for distributed calculation.

threads

The number of threads available locally for execution, 1 by default.

gpu_id

The identifier of the GPU to use, None by default. Must be None, or a non-negative integer.

mpi

The MPI communicator (see mpi_comm). mpi must be None, or an MPI communicator.

Contexts can be queried for information about which features a context has enabled, whether it has a GPU, how many threads are in its thread pool.

has_gpu

Query whether the context has a GPU.

has_mpi

Query whether the context uses MPI for distributed communication.

threads

Query the number of threads in the context’s thread pool.

ranks

Query the number of distributed domains. If the context has an MPI communicator, return is equivalent to MPI_Comm_size. If the communicator has no MPI, returns 1.

rank

The numeric id of the local domain. If the context has an MPI communicator, return is equivalent to MPI_Comm_rank. If the communicator has no MPI, returns 0.

Here are some simple examples of how to create a context:

import arbor
import mpi4py.MPI as mpi

# Construct a context that uses 1 thread and no GPU or MPI.
context = arbor.context()

# Construct a context that:
#  * uses 8 threads in its thread pool;
#  * does not use a GPU, reguardless of whether one is available
#  * does not use MPI.
alloc   = arbor.proc_allocation(8, None)
context = arbor.context(alloc)

# Construct a context that uses:
#  * 4 threads and the first GPU;
#  * MPI_COMM_WORLD for distributed computation.
alloc   = arbor.proc_allocation(4, 0)
comm    = arbor.mpi_comm(mpi.COMM_WORLD)
context = arbor.context(alloc, comm)