Help running parallel copies of R code that calls a Fortran DLL

[woodward]

Hi Rsters! I have a legacy simulation model written in Fortran 95 that I am am calibrating using the excellent BayesianTools package. I was running it from R using the .Fortran() function, but was having memory problems for long runs/chains because this only works under 32-bit R with 32-bit DLLs.

Recently I solved this by compiling the Fortran inside an R package following the excellent instructions at
https://www.avrahamadler.com/2018/12/09/the-need-for-speed-part-1-building-an-r-package-with-fortran/

This allowed the calibration to run under 64-bit R, solving the memory issues, but didn't really make it faster.

I am exploring options for speeding it up via parallelisation. I am running 3 independent calibration chains (each has 3 internal chains in DREAMzs) that should be easy to run on separate cores (I have an Intel Core i7 with 8 logical processors and 24Gb RAM running Windows 10).

I have it running using the parallel package and parLapply(), which uses more CPU % than the serial version, but elapsed time is the same. Is there some bottleneck caused by calling a Fortran DLL? Even though it is compiled in an R package? Is there only a single instance of the DLL or can I clone it to all the nodes?

Thanks!

PS: I'll see if I can come up with a meaningful reprex... hmm, it works for the SimpFort package vignette from the link above (see below), so maybe Fortran is not the problem...

library(SimpFort) # package that uses Fortran DLL
r <- 3 # number of times to run
n <- 100000000 # size of calculation
x <- rep(100000,n)

print("Run serial")
#> [1] "Run serial"
print(system.time(
  print(unlist(lapply(X=1:r, FUN=function(X){
    LLC_f(x, 50000, 90000)
  })))
))
#> [1] 1e+12 1e+12 1e+12
#>    user  system elapsed 
#>    0.35    0.00    0.36

library(parallel)
cl <- makeCluster(r)
clusterEvalQ(cl, library(SimpFort))
#> [[1]]
#> [1] "SimpFort"  "stats"     "graphics"  "grDevices" "utils"     "datasets" 
#> [7] "methods"   "base"     
#> 
#> [[2]]
#> [1] "SimpFort"  "stats"     "graphics"  "grDevices" "utils"     "datasets" 
#> [7] "methods"   "base"     
#> 
#> [[3]]
#> [1] "SimpFort"  "stats"     "graphics"  "grDevices" "utils"     "datasets" 
#> [7] "methods"   "base"
clusterExport(cl, "x")

print("Run parallel")
#> [1] "Run parallel"
print(system.time(
  print(unlist(parLapply(cl, X=1:r, fun=function(X){
    LLC_f(x, 50000, 90000)
  })))
))
#> [1] 1e+12 1e+12 1e+12
#>    user  system elapsed 
#>    0.00    0.00    0.14

^{Created on 2019-04-10 by the reprex package (v0.2.1)}

[woodward]

Solved it! Turns out I was timing the code wrongly, and triple counting the time spend in the parallel version!

[andresrcs]

If your question's been answered (even by you!), would you mind choosing a solution? It helps other people see which questions still need help, or find solutions if they have similar problems. Here’s how to do it:

FAQ: How do I mark a solution?

If your question has been answered, don't forget to mark the solution! How do I mark a solution? Find the reply you want to mark as the solution and look for the row of small gray icons at the bottom of that reply. Click the one that looks like a box with a checkmark in it: [image] Hovering over the mark solution button shows the label, "Select if this reply solves the problem". If you don't see the mark solution button, try clicking the three dots button ( ••• ) to expand the full set of options. When a solution is chosen, the icon turns green and the hover label changes to: "Unselect if this reply no longer solves the problem". Success! [solution_reply_author] …