Dear Professor Ripley,
PkgB does not need to be on the search path---importing into the namespace
is fine. I did not realize that namespace scoping ensured that if a
cluster is created from within a package, that packages entire environment
tree is available on all the workers.
I tried to apply how makeCluster works from an interactive R session, where
functions from packages that are loaded when the cluster is created are not
available on the workers, to how it works from within a package.
Thanks for your reply,
Josh
On Fri, Aug 8, 2014 at 4:47 PM, Prof Brian Ripley <ripley at stats.ox.ac.uk>
wrote:
The safe, elegant way to do this is to use namespace scoping: it is still
not at all clear why 'other code' needs PkgB *on the search path*.
In other cases seen in CRAN submissions, 'other code' has been in PkgA's
namespace, and hence things in PkgB's exports have been visible as it was
imported by PkgA and hence in the environment tree for functions in PkgA.
Then namespace scoping will ensure that PkgB's namespace is loaded on the
cluster workers.
On 08/08/2014 00:58, Joshua Wiley wrote:
Someone kindly pointed out that it is not clear from my email why Depends
will not work. A more complete example is:
PkgA:
f <- function(ncores) {
cl <- makeCluster(ncores)
clusterEvalQ(cl, {
require(PkgB)
})
[other code]
### this is the code I want to work and need to be able to call
### PkgB functions on each of the cluster slaves
output <- parLapply(cl, 1:n, function(i) {
[code from my package and using some functions from PkgB]
})
}
As far as I know, just because I add PkgB to the Depends (or imports,
whatever) of PkgA, does not mean that the cluster started by PkgA will
automatically have PkgB loaded and functions available.
Thanks!
On Fri, Aug 8, 2014 at 9:35 AM, Joshua Wiley <jwiley.psych at gmail.com>
wrote:
Dear All,
What is the preferred way for Package A, to initialize a cluster, and
load
Package B on all nodes?
I am writing a package that parallelizes some functions through the use
of
a cluster if useRs are on a Windows machine (using parLapply and family).
I also make use of another package in some of my code, so it is
necessary
to load the required packages on each slave once the cluster is started.
Right now, I have done this, by evaluating require(packages) on each
slave; however, Rcmd check has a note that I should remove the "require"
in
my code.
Thanks!
Josh
--
Joshua F. Wiley
Ph.D. Student, UCLA Department of Psychology
http://joshuawiley.com/
Senior Analyst, Elkhart Group Ltd.
http://elkhartgroup.com
Office: 260.673.5518
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595