R help - Feb 2007 - Snow vs Rmpi

Hi, 

I have few high-level questions about the Snow and Rmpi packages . I understand
that Snow uses Rmpi as one of possible transport layers, yet my questions about
user experience, not technical details:

1. Does Snow install and work well in Windows? 
2. Interruptibility. I understand that currently it is impossible to interrupt a
running top-level command in Snow ( Ctl-c or the likes), the only way to kill
slave processes is to kill the master R process. Is this accurate? What about
Rmpi ? Is there any difference between Windows and Linux?
3. When the master process dies , is it guaranteed that the slaves will die too?
How reliable is this (I've seen some applications, not related to R, that
were flaky about killing slaves)

Thank you very much for your help, 
Vadim 



	[[alternative HTML version deleted]]

Dear Vadim,

On 2/14/07, Vadim Ogranovich <vogranovich at jumptrading.com>
wrote:
> Hi,
>
> I have few high-level questions about the Snow and Rmpi packages . I
understand that Snow uses Rmpi as one of possible transport layers, yet my
questions about user experience, not technical details:
>
> 1. Does Snow install and work well in Windows?
> 2. Interruptibility. I understand that currently it is impossible to
interrupt a running top-level command in Snow ( Ctl-c or the likes), the only
way to kill slave processes is to kill the master R process. Is this accurate?
What about Rmpi ? Is there any difference between Windows and Linux?

I've never used any of those under Windoze. I think your statement is
accurate under Linux. (In fact, I often get rid of any of those Rmpis
gone astray by issuing a lamhalt and/or lamwipe).
> 3. When the master process dies , is it guaranteed that the slaves will die
too? How reliable is this (I've seen some applications, not related to R,
that were flaky about killing slaves)

If you use an orderly exit procedure (mpi.close.Rslaves(); mpi.quit())
I've never, ever, seen badly behaved Rmpi slaves. But I've seen them
under strange circumstances (I think network problems that messed up
the lam universe ?).

A kind of fail proof approach, if you can afford it, is to use
different lam universes (using the LAM_MPI_SESSION_SUFFIX) for
different simultaneous runs. Then, if one particular run behaves
poorly, you can issue a lamhalt/lamwipe for just that LAM universe.

A final suggestion: you might want to take a look at the papply
package, which does load-balancing and allows you to run sequential
(if there is no lam universe), and thus makes debugging much simpler.

R.

>
> Thank you very much for your help,
> Vadim
>
>
>
>         [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

-- 
Ramon Diaz-Uriarte
Statistical Computing Team
Structural Biology and Biocomputing Programme
Spanish National Cancer Centre (CNIO)
http://ligarto.org/rdiaz