[R] Connecting to "heterogenous" cluster using makeSOCKcluster of SNOW-package

Melwin win at comets.de
Tue Jun 16 12:18:04 CEST 2009


Thanks to Luke Tiernay and some experimenting I found out some issues. I  
don't claim this infomation is complete, but it may be helpful for anyone  
experimenting with SNOW on Linux:

- environment variables PATH and R_SNOW_LIB need to be set on master and  
slaves.
(manually or permanent in ~/.bashrc (adjust your pathnames accordingly).  
Don't use ~ for home directory, instead specify the full path. Example:

PATH=$PATH:/home/johnson/R/x86_64-unknown-linux-gnu-library/2.8/snow
export PATH

R_SNOW_LIB="/home/johnson/R/x86_64-unknown-linux-gnu-library/2.8"
export R_SNOW_LIB

- when creating a cluster, as in the example below, the "snowlib" argument  
seems ignored, so use the environment variable as stated above:
	lnxOptions <- list(host = "GEOECOLOGY\\\\johnson at geco01",snowlib =  
"/home/GEOECOLOGY/johnson/R/i686-suse-linux-gnu-library/2.8")
	cl <- makeCluster(rep(list(lnxOptions), 2), type = "SOCK")

- when specifying the hostname, most likely you'll need to include your  
username (see example above). If this user is part of a domaine (e.g.  
GEOECOLOGY), this prepends the username separated by quadruple backslashes  
(see above)

- Finally, various postings and personal experience suggest that the  
"SOCK"-based connection tends to break down easily, making it only  
feasible to be used on a single multi-core computer. MPI and PVM seem more  
tested, but are virtually impossible to install for Linux-newbies,  
especially under Suse-Linux.

I begin to wonder what "parallelizing the work" really means.




More information about the R-help mailing list