ergm-parallel: Parallel Processing in the 'ergm' Package

In ergm: Fit, Simulate and Diagnose Exponential-Family Models for Networks

Parallel Processing in the ergm Package

Description

Using clusters multiple CPUs or CPU cores to speed up ERGM estimation and simulation.

The ergm.getCluster function is usually called internally by the ergm process (in ergm_MCMC_sample()) and will attempt to start the appropriate type of cluster indicated by the control.ergm() settings. It will also check that the same version of ergm is installed on each node.

The ergm.stopCluster shuts down a cluster, but only if ergm.getCluster was responsible for starting it.

The ergm.restartCluster restarts and returns a cluster, but only if ergm.getCluster was responsible for starting it.

nthreads is a simple generic to obtain the number of parallel processes represented by its argument, keeping in mind that having no cluster (e.g., NULL) represents one thread.

Usage

ergm.getCluster(control = NULL, verbose = FALSE, stop_on_exit = parent.frame())

ergm.stopCluster(..., verbose = FALSE)

ergm.restartCluster(control = NULL, verbose = FALSE)

set.MT_terms(n)

get.MT_terms()

nthreads(clinfo = NULL, ...)

## S3 method for class 'cluster'
nthreads(clinfo = NULL, ...)

## S3 method for class ''NULL''
nthreads(clinfo = NULL, ...)

## S3 method for class 'control.list'
nthreads(clinfo = NULL, ...)

Arguments

control	a control.ergm() (or similar) list of parameter values from which the parallel settings should be read; can also be NULL, in which case an existing cluster is used if started, or no cluster otherwise.
verbose	A logical or an integer to control the amount of progress and diagnostic information to be printed. FALSE/0 produces minimal output, with higher values producing more detail. Note that very high values (5+) may significantly slow down processing.
stop_on_exit	An environment or NULL. If an environment, defaulting to that of the calling function, the cluster will be stopped when the calling the frame in question exits.
...	not currently used
n	an integer specifying the number of threads to use; 0 (the starting value) disables multithreading, and -1 or NA sets it to the number of CPUs detected.
clinfo	a cluster or another object.

Details

For estimation that require MCMC, ergm can take advantage of multiple CPUs or CPU cores on the system on which it runs, as well as computing clusters through one of two mechanisms:

Running MCMC chains in parallel

Packages parallel and snow are used to to facilitate this, all cluster types that they support are supported.

The number of nodes used and the parallel API are controlled using the parallel and parallel.type arguments passed to the control functions, such as control.ergm().

The ergm.getCluster() function is usually called internally by the ergm process (in ergm_MCMC_sample()) and will attempt to start the appropriate type of cluster indicated by the control.ergm() settings. The ergm.stopCluster() is helpful if the user has directly created a cluster.

Further details on the various cluster types are included below.

Multithreaded evaluation of model terms

Rather than running multiple MCMC chains, it is possible to attempt to accelerate sampling by evaluating qualified terms' change statistics in multiple threads run in parallel. This is done using the OpenMP API.

However, this introduces a nontrivial amont of computational overhead. See below for a list of the major factors affecting whether it is worthwhile.

Generally, the two approaches should not be used at the same time without caution. In particular, by default, cluster slave nodes will not “inherit” the multithreading setting; but ⁠parallel.inherit.MT=⁠ control parameter can override that. Their relative advantages and disadvantages are as follows:

• Multithreading terms cannot take advantage of clusters but only of CPUs and cores.

• Parallel MCMC chains produce several independent chains; multithreading still only produces one.

• Multithreading terms actually accellerates sampling, including the burn-in phase; parallel MCMC's multiple burn-in runs are effectively “wasted”.

Value

set.MT_terms() returns the previous setting, invisibly.

get.MT_terms() returns the current setting.

Different types of clusters

PSOCK clusters

The parallel package is used with PSOCK clusters by default, to utilize multiple cores on a system. The number of cores on a system can be determined with the detectCores() function.

This method works with the base installation of R on all platforms, and does not require additional software.

For more advanced applications, such as clusters that span multiple machines on a network, the clusters can be initialized manually, and passed into ergm() and others using the parallel control argument. See the second example below.

MPI clusters

To use MPI to accelerate ERGM sampling, pass the control parameter parallel.type="MPI". ergm requires the snow and Rmpi packages to communicate with an MPI cluster.

Using MPI clusters requires the system to have an existing MPI installation. See the MPI documentation for your particular platform for instructions.

To use ergm() across multiple machines in a high performance computing environment, see the section "User initiated clusters" below.

User initiated clusters

A cluster can be passed into ergm() with the parallel control parameter. ergm() will detect the number of nodes in the cluster, and use all of them for MCMC sampling. This method is flexible: it will accept any cluster type that is compatible with snow or parallel packages.

When is multithreading terms worthwhile?

• The more terms with statistics the model has, the more benefit from parallel execution.

• The more expensive the terms in the model are, the more benefit from parallel execution. For example, models with terms like gwdsp will generally get more benefit than models where all terms are dyad-independent.

• Sampling more dense networks will generally get more benefit than sparse networks. Network size has little, if any, effect.

• More CPUs/cores usually give greater speed-up, but only up to a point, because the amount of overhead grows with the number of threads; it is often better to “batch” the terms into a smaller number of threads than possible.

• Any other workload on the system will have a more severe effect on multithreaded execution. In particular, do not run more threads than CPUs/cores that you want to allocate to the tasks.

• Under Windows, even compiling with OpenMP appears to introduce unacceptable amounts of overhead, so it is disabled for Windows at compile time. To enable, delete src/Makevars.win and recompile from scratch.

Note

The this is a setting global to the ergm package and all of its C functions, including when called from other packages via the Linking-To mechanism.

Examples

# Uses 2 SOCK clusters for MCMLE estimation
data(faux.mesa.high)
nw <- faux.mesa.high
fauxmodel.01 <- ergm(nw ~ edges + isolates + gwesp(0.2, fixed=TRUE), 
                     control=control.ergm(parallel=2, parallel.type="PSOCK"))
summary(fauxmodel.01)

ergm documentation built on Oct. 7, 2024, 5:08 p.m.