the random data generating function. See the details
section
rand.arg
additional argument list to rand.gen.
statistic
the statistic function to be simulated. See the
details section
nsim
the number of simulation carried on a slave which is
counted as one slave job.
run
the number of looping. See the details section.
slaveinfo
if TRUE, the numbers of jobs finished by slaves
will be displayed.
sim.seq
if reproducing the same simulation is desirable, set it
to the integer vector .mpi.parSim generated in previous simulation.
simplify
logical; should the result be simplified to a vector or
matrix if possible?
comm
a communicator number
...
optional arguments to statistic
Details
It is assumed that one simulation is carried out as
statistic(rand.gen(n)), where rand.gen(n) can return any
values as long as statistic can take them. Additional arguments can
be passed to rand.gen by rand.arg as a list. Optional
arguments can also be passed to statistic by the argument
....
Each slave job consists of replicate(nsim,statistic(rand.gen(n))),
i.e., each job runs nsim number of simulation. The returned values
are transported from slaves to master.
The total number of simulation (TNS) is calculated as follows. Let
slave.num be the total number of slaves in a comm and it is
mpi.comm.size(comm)-1. Then TNS=slave.num*nsim*run and the total
number of slave jobs is slave.num*run, where run is the number of
looping from master perspective. If run=1, each slave will run one slave
job. If run=2, each slave will run two slaves jobs on average, and so on.
The purpose of using run has two folds. It allows a tuneup
of slave job size and total number of slave jobs to deal with two
different cluster environments. On a cluster of slaves with equal CPU
power, run=1 is often enough. But if nsim is too big, one
can set run=2 and the slave jog size to be nsim/2 so that
TNS=slave.num*(nsim/2)*(2*run). This may improve R computation
efficiency slightly. On a cluster of slaves with different CPU power, one
can choose a big value of run and a small value of nsim
so that master can dispatch more jobs to slaves who run faster than
others. This will keep all slaves busy so that load balancing is
achieved.
The sequence of slaves who deliver results to master are saved into
.mpi.parSim. It keeps track which part of results done by which slaves.
.mpi.parSim can be used to reproduce the same simulation result if the same
seed is used and the argument sim.seq is equal to .mpi.parSim.
See the warning section before you use mpi.parSim.
Value
The returned values depend on values returned by replicate
of statistic(rand.gen(n)) and the total number of simulation
(TNS). If statistic returns a single value, then the result is a
vector of length TNS. If statistic returns a vector (list) of
length nrow, then the result is a matrix of dimension
c(nrow, TNS).
Warning
It is assumed that a parallel RNG is used on all slaves. Run
mpi.setup.rngstream on the master to set up a parallel RNG. Though mpi.parSim
works without a parallel RNG, the quality of simulation is not guarantied.
mpi.parSim will automatically transfer rand.gen
and statistic to slaves. However, any functions that
rand.gen and statistic reply on but are not on slaves
must be transfered to slaves before using mpi.parSim. You
can use mpi.bcast.Robj2slave for that purpose. The same is
applied to required packages or C/Fortran codes. You can use either
mpi.bcast.cmd or put required(package) and/or
dyn.load(so.lib) into rand.gen and statistic.
If simplify is TRUE, sapply style simplication is applied. Otherwise a list of length
slave.num*run is returned.