4

I am trying to teach myself how to use the Snowfall package, and I have run into the following problem when I try a function that calls a second function (this is a simplified use case of what I ultimately want to implement).

I currently have:

library (snowfall)
f1 <- function(n) { return (n-1) }
f2 <- function(n) { return (f1(n)^2) }
# initialize cluster
sfInit (parallel=TRUE , cpus=4)
# parallel computing
result <- sfLapply(1:10, f2)
# stop cluster
sfStop ()

but I receive the error message:

Error in checkForRemoteErrors(val) :
  4 nodes produced errors; first error: could not find function "f1"

However, if I then run lapply(1:10, f2) I receive the following output:

lapply(1:10, f2)
[[1]]
[1] 0

[[2]]
[1] 1

[[3]]
[1] 4

[[4]]
[1] 9

[[5]]
[1] 16

[[6]]
[1] 25

[[7]]
[1] 36

[[8]]
[1] 49

[[9]]
[1] 64

[[10]]
[1] 81

I ultimately want to use snowfall to implement a parallelized search procedures for multidimensional minimization problems, so will definitely need to be able to call functions from the main parallelized function.

Can anyone help with this?

billelev
  • 369
  • 2
  • 13

2 Answers2

4

You need to export the f1 function to the workers using the sfExport function between sfInit and sfLapply:

sfExport('f1')

This is the snowfall equivalent to the snow clusterExport function.

To export multiple variables, you can either use multiple arguments or the list argument:

sfExport('f1', 'x')
sfExport(list=c('f1', 'x'))

To export all variables in your global environment, use sfExportAll:

sfExportAll()
Steve Weston
  • 19,197
  • 4
  • 59
  • 75
  • Perfect...thank you. And for complicated implementations, presumably sfExportAll() would work? – billelev Sep 09 '13 at 19:02
  • @user2762408 `sfExportAll` is convenient if you want to export everything in your global environment, and the `except` option can be used to exclude variables that are not wanted. You can also export multiple variables with `sfExport`, but you do have to name them all. – Steve Weston Sep 09 '13 at 19:11
0

I do not know the specifics of the snowfall package, but I do know that when doing parallel computing in R, you need to assume that each core you pass information to is a brand-new R instance. Meaning, regardless of what you pass through your host prior to registering clusters, there is just the vanilla instance of R on your clones.

It looks like there is a call sfExportAll() which will put all of your global variables onto the clone instances. I would try it out, but I can't run snowfall on my Windows machine.

dylanjf
  • 191
  • 1
  • 4