Parallel simulations

Simulations are necessary to compute the confidence envelopes of the main functions of the package. They can be run in parallel thanks to the future and doFuture (Bengtsson 2021) packages.

Example

Estimating the confidence envelope of the M function under the null hypothesis of random labeling of the point types relies on the MEnvelope() function. Simulations are run sequentially. The computing time can be measured.

library("dbmss")
system.time(
  MEnvelope(
    paracou16, 
    r = NULL, 
    NumberOfSimulations = 5000, 
    Alpha = 0.05, 
    ReferenceType = "V. Americana", 
    NeighborType = "Q. Rosea", 
    SimulationType = "RandomLabeling", 
    Global = TRUE
  )  
)

##    user  system elapsed 
##  48.267   1.258  22.807

It can be parallelized by setting a hardware strategy. The following example launches several simultaneous copies of R on the local computer: it works on all operating systems. The number of parallel processes here is the number of logical CPUs.

library("doFuture")
plan(multisession, workers = availableCores())

The same set of simulations is now run faster with argument parallel = TRUE.

system.time(
  MEnvelope(
    paracou16, 
    r = NULL, 
    NumberOfSimulations = 5000, 
    Alpha = 0.05, 
    ReferenceType = "V. Americana", 
    NeighborType = "Q. Rosea", 
    SimulationType = "RandomLabeling", 
    Global = TRUE, 
    parallel = TRUE
  )
)

##    user  system elapsed 
##   1.443   0.080  21.090

The default strategy is sequential. It should be restored.

plan(sequential)

Other possible strategies are forking (multicore), unavailable on Windows, and cluster, on several computers.

See the documentation of the future package, namely future::plan().

Progress bar

Parallel simulations can report their progress thanks to the progressr package. They ignore the verbose argument but rely on the parallel_pgb_refresh argument to refresh the progress bar more or less often.

The progress bar must be activated and its interface chosen, for example by the following code, to be run before calling the MEnvelope() function:

library("progressr")
handlers(global = TRUE)
handlers("txtprogressbar")

Note that the choice of the progress bar is free: see ?progressr.

To limit the overhead of updating the progress bar, this can be done during a fraction of the simulations. parallel_pgb_refresh is $1/10$ by default, meaning that the progress bar display is updated one simulation out of ten. This fraction can be increased to 1 if each simulation takes a long time (say, one second or more) or decreased to improve the computing performance. In the example above, $1/100$ is a reasonable choice on a laptop computer.

Limits

Setting up the parallel code and gathering the results takes time so the performance of sequential simulations may be better if they are fast.

References

Bengtsson, Henrik. 2021. “A Unifying Framework for Parallel and Distributed Processing in R Using Futures.” The R Journal 13 (2): 208. https://doi.org/10.32614/RJ-2021-048.