[R] snow documentation comments

[Luke Tierney]

On Wed, 24 Mar 2004, Ross Boylan wrote:

> On Wed, 2004-03-24 at 08:03, Luke Tierney wrote:
> > On Tue, 23 Mar 2004, Ross Boylan wrote:
> > 
> > > There are a few points I found unclear or unmentioned in the snow
> > > documentation (mostly I looked at the cluster.html web page).  I thought
> > > I'd mention them here.
> > > 
> > > What is the start up environment for the children?
> > > --------------------------------------------------
> > > My best guess at the answer is in parentheses
> > > Do they inherit shell variables? (no)
> > > Do they inherit variables set in R or other aspects of the R
> > > environment? (no)
> > > What directory does it start in? (the directory you are running in)
> > > What user are you? (same as original)
> > > 
> > > I realize some of these answers might depend on the parallelization
> > > layer you are using, or whether you select the homogenous option.
> > 
> > They do
> I've since realized they may also depend on the exact version of the
> package used (e.g., newer lams apparently have the ability to export
> environment variables) and the way it is invoked.
> 
> Personally, I'd even find a single sentence saying explicitly that "it
> depends" clarifying.  It would probably also be good to discuss the
> behavior of the socket-based communication, since that it supplied by
> the snow package.  Information on the other interfaces would be nice
> too, but probably hazardous to provide accurately.
> 
> I'll not for the record that my answers above are using Rmpi with
> lam/mpi 6.5.8-2 on Debian, running just on my local computer.  And some
> of the answers are speculative!

I'll make a note to add a sentence.

Initial configuration is definitely the harders part about using snow,
especially as it by nature has to depend on aspects of the local
infrastructure.  Working on both the documentation and on some tools
to handle a richer range of scenarios easily is definitely something
we need to work on.

> > 
> > > Number of nodes or number of children?
> > > --------------------------------------
> > > I thought all the counts of nodes (e.g., in makeCluster(5)) were of the
> > > total nodes in the cluster, i.e., children + 1.  However, I did a
> > > makeCluster(2) and I got 2 additional R processes running.  Have I
> > > misunderstood the semantics, or is it essentially an implementation
> > > detail that the master node starts a new R process?
> > 
> > makeCluster(2) creates a cluster of two processes that the master
> > process uses.  So there are a total of three processes.

> So work that is distributed via snow goes to the two processes only?

That is correct--the master just waits intil somethng comes back from
at least one of the worker processes.

luke


-- 
Luke Tierney
University of Iowa                  Phone:             319-335-3386
Department of Statistics and        Fax:               319-335-3017
   Actuarial Science
241 Schaeffer Hall                  email:      luke at stat.uiowa.edu
Iowa City, IA 52242                 WWW:  http://www.stat.uiowa.edu