[R-SIG-Finance] Running R as a server or in a cluster
Brian G. Peterson
brian at braverock.com
Wed Sep 26 16:53:22 CEST 2007
Joshua Reich wrote:
> Yes - its JBOC (just a bunch of computers). You provide them with a disk
> image (of sorts) and they will load it on to as many computers as you
> request. Images are loaded and machines are requested via a web services
> API. Initially you can request up to 20 machines - but if you email them
> you can ask for more. All network bandwidth between machines is free,
> but there is a per GB transfer charge for external connectivity - I
> can't recall what the rate is, but it is very reasonable.
> Not being a specialized grid environment, all inter-node communication
> and scheduling has to be handled by your own application. But for the
> price, that's not too bad.
> While I was aware of SNOW, I'm not familiar with the other clustering
> approaches mentioned earlier in this thread. What special sauce does Sun
> provide to make running on a grid easier than running on a JBOC style
Sun and HP both contributed to the development of Parallel-R, I believe.
So I would assume that the sun cluster provides these capabilities.
One simple approach for highly-parallizable calculations that I've seen
has been to use a parallel version of the apply function.
I think that a cluster-aware portfolio optimization package framework
should be relatively straightforward to put together in R. Other
analyses would need to be taken on a case-by-case basis.
Rserve can work well in front of a cluster environment to run individual
self-contained queries without having to program for a cluster
environment. Other analysis would/might require that your code be
cluster-aware, and send out and collect distributed jobs.
More information about the R-SIG-Finance