On Mon, 7 Oct 2002, Vijay Mani wrote:
> I am not sure where I need to specify the tcp or sysv or usysv
> RPI when installing lam, and is it supported by 6.5.6?
> I was also wondering if anyone has managed to write any code( no matter how
> small) that runs better on dual processor machines than twice the number of
> single processor machines.
All three rpis listed above (tcp, usysv, and sysv) are fully supported in
the 6.5.6 release of LAM/MPI. You can use any of the three and it should
"just work". Unless you are running on Mac OS X 10.0 or 10.1, in which
case you only have tcp. But that probably isn't a problem for you :).
Yes, it is possible (sometimes even easy) to write code that runs faster
on a cluster of n dual-cpu machines than 2*n single-cpu machines.
However, there are some gotchas. Your memory requirements per-box double
in the SMP case, so you have to get more RAM per machine (same per CPU) if
you have memory-intensive jobs. Your memory bandwidth also increases
greatly. In addition to normal application memory usage, the memory bus
is now being used as a transport agent by the MPI. Slow memory might
cause contention issues on the bus.
As with most things in parallel computing, a lot depends on your
application. If you can find a vendor willing to let you test your
application, that would be the best idea. But in general, that isn't
going to happen. Also, if you go with SMP machines, you should try all
three rpis - each application behaves differently under each rpi. Even on
SMP machines, tcp can occasionally be the best performing rpi.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
_______________________________________________
This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
