doMPI package provides a parallel backend for the foreach
package. It is similar to the doSNOW package, but uses Rmpi
directly. This allows it to do more, and execute more efficiently.
It can also make use of the multicore package to execute tasks
across multiple cores on the worker nodes. This is can give very good
performance on a computer cluster with multicore processors.doMPI. They are specified to foreach as a list using the
.options.mpi argument. The currently supported options are:
chunkSize |
Number of tasks to send at a time to the cluster workers |
info |
Display extra information, particularly about exported variables |
initEnvir |
A function to be called on each worker before executing any tasks |
initArgs |
List of extra arguments to pass to the initEnvir function |
initEnvirMaster |
A function called on the master at the same time as initEnvir |
initArgsMaster |
List of extra arguments to pass to the initEnvirMaster function |
finalEnvir |
A function to be called on each worker after executing all tasks |
finalArgs |
List of extra arguments to pass to the finalEnvir function |
profile |
Display profiling information from the master's point of view |
bcastThreshold |
Used to decide whether to piggy-back or broadcast job data |
forcePiggyback |
Always piggy-back job environment with first task to each worker |
nocompile |
Don't compile the R expression |
seed |
Starting seed for tasks |
The chunkSize option is particularly important, since it can be
much more efficient to send more than one task at a time to the workers,
particularly when the tasks execute quickly. Also, it can allow the
workers to execute those tasks in parallel using the mclapply
function from the multicore package. The default value is
1.
The info option is used to print general information that is
specific to the doMPI backend. This includes information on what
variables are exported, for example. The default value is FALSE.
The initEnvir option is useful for preparing the workers to
execute the subsequent tasks. The execution environment is passed as
the first argument to this function. That allows you to define new
variables in the environment, for example. If initArgs is
defined, the contents of the list will be passed as arguments to the
initEnvir function after the environment object.
The initEnvirMaster option is useful if you want to send data
from the master to the workers explicitly, perhaps using
mpi.bcast. This avoids object serialization, which could improve
performance for large matrices, for example. The initArgsMaster
option works like initArgs, however, it is probably less useful,
since the initEnvirMaster function runs locally, and can access
variables via lexical scoping.
The finalEnvir option is useful for “finalizing” the execution
environment. It works pretty much the same as the initEnvir
function, getting extra arguments from a list specified with the
finalArgs option.
The profile option is used to print profiling information at the
end of the %dopar% execution. It basically lists the time spent
sending tasks to the workers and retrieving results from them. The
default value is FALSE.
The bcastThreshold option is used to decide whether to piggy-back
the job data, or broadcast it. The job data is serialized, and if it is
smaller than bcastThreshold, it is piggy-backed, otherwise, it is
broadcast. Note that if you want to force piggy-backing, you should use
the forcePiggyback, rather than setting bcastThreshold to
a very large value. That avoids serializing the job data twice, which
can be time consuming.
The forcePiggyback option is used to force the job data to be
“piggy-backed” with the first task to each of the workers. If the
value is FALSE, the data may still be piggy-backed, but it is not
guaranteed. In general, the job data is only piggy-backed if it is
relatively small. The default value is FALSE.
The nocompile option is used to disable compilation of the
R expression in the body of the foreach loop. The default value is
FALSE.
The seed option is used for achieving reproducible results. If
set to a single numeric value, such as 27, it is converted to a
value that can be passed to the nextRNGSubStream function from
the parallel package. This value is assigned to the global
.Random.seed variable on some cluster worker when it executes the
first task (or task chunk). The nextRNGSubStream function is
used to generate subsequent values that are assigned to
.Random.seed when executing subsequent tasks. Thus, RNG
substreams are associated with tasks, rather than workers.
This is necessary for reproducible results, since the doMPI
package uses load balancing techniques that can result in different
tasks being executed by different workers on different runs of the same
foreach loop. The default value of the seed option is
NULL.
Additional documentation is available on the following functions:
startMPIcluster |
Create and start an MPI cluster object |
registerDoMPI |
Register a cluster object to be used with %dopar% |
closeCluster |
Shutdown and close a cluster object |
clusterSize |
Return the number of workers associated with a cluster object |
setRngDoMPI |
Initialize parallel random number generation on a cluster |
For a complete list of functions with individual help pages,
use library(help="doMPI").
Use the command vignette("doMPI") to view the vignette entitled
“Introduction to doMPI”.
Also, there are a number of doMPI example scripts in the
examples directory of the doMPI installation.