doParallel issue with inline function on Windows 7 (works on Linux)

Question

I am using R 3.0.1 both on Windows 7 and Linux (SUSE Server 11 (x86_64)). The following example code produces an error on Windows but not on Linux. All the toolboxes listed are up-to-date in both machines. The Windows error is:

Error in { : task 1 failed - "NULL value passed as symbol address"

If I change %dopar% to %do%, the Windows code runs without any errors. My initial guess was that this relates to some configuration issue in Windows and I tried reinstalling Rcpp and R but that did not help. The error seems to be related to scoping - if I define and compile the function cFunc inside f1, then %dopar% works but, as expected, it is very slow since we are calling the compiler once for each task.

Does anyone have some insights on why the error happens or suggestions on how to fix it?

library(inline)
sigFunc <- signature(x="numeric", size_x="numeric")
code <- ' double tot =0;
for(int k = 0; k < INTEGER(size_x)[0]; k++){
tot += REAL(x)[k];
};
return ScalarReal(tot);
' 
cFunc <- cxxfunction(sigFunc, code)

f1 <- function(){
x <- rnorm(100)
a <- cFunc(x=x, size_x=as.integer(length(x)))
return(a)
}

library(foreach)
library(doParallel)
registerDoParallel()
# this produces an error in Windows but not in Linux
res <- foreach(counter=(1:100)) %dopar% {f1()}
# this works for both Windows and Linux
res <- foreach(counter=(1:100)) %do% {f1()}

# The following is not a practical solution, but I can compile cFunc inside f1 and then     this works in Windows but it is very slow
f1 <- function(){
library(inline)
sigFunc <- signature(x="numeric", size_x="numeric")

code <- ' double tot =0;
for(int k = 0; k < INTEGER(size_x)[0]; k++){
tot += REAL(x)[k];
};
return ScalarReal(tot);
' 
cFunc <- cxxfunction(sigFunc, code)
x <- rnorm(100)
a <- cFunc(x=x, size_x=as.integer(length(x)))
return(a)
}
# this now works in Windows but is very slow
res <- foreach(counter=(1:100)) %dopar% {f1()}

Thanks! Gustavo

Answer 1

The error message "NULL value passed as symbol address" is unusual, and isn't due to the function not being exported to the workers. The cFunc function just doesn't work after being serialized, sent to a worker, and unserialized. It also doesn't work when it's loaded from a saved workspace, which results in the same error message. That doesn't surprise me much, and it may be a documented behavior of the inline package.

As you've demonstrated, you can work-around the problem by creating cFunc on the workers. To do that efficiently, you need to do it only once on each of the workers. To do that with the doParallel backend, I would define a worker initialization function, and execute it on each of the workers using the clusterCall function:

worker.init <- function() {
  library(inline)
  sigFunc <- signature(x="numeric", size_x="numeric")
  code <- ' double tot =0;
  for(int k = 0; k < INTEGER(size_x)[0]; k++){
  tot += REAL(x)[k];
  };
  return ScalarReal(tot);
  '
  assign('cFunc', cxxfunction(sigFunc, code), .GlobalEnv)
  NULL
}

f1 <- function(){
  x <- rnorm(100)
  a <- cFunc(x=x, size_x=as.integer(length(x)))
  return(a)
}

library(foreach)
library(doParallel)
cl <- makePSOCKcluster(3)
clusterCall(cl, worker.init)
registerDoParallel(cl)
res <- foreach(counter=1:100) %dopar% f1()

Note that you must create the PSOCK cluster object explicitly in order to call clusterCall.

The reason that your example worked on Linux is that the mclapply function is used when you call registerDoParallel without an argument, while on Windows a cluster object is created and the clusterApplyLB function is used. Functions and variables aren't serialized and sent to the workers when using mclapply, so there is no error.

It would be nice if doParallel included support for initializing the workers without the need for using clusterCall, but it doesn't yet.