Scyld and pvfs ...

[Daniel Ridge]

Franz,

On Tue, 14 Nov 2000, Franz Marini wrote:

> 4) (quite a long list of questions ...) we're using the CPMD
> (Carr-Parrinello Molecur Dynamics) program on our current beowulf, but we
> have some performance problems w/ mpich 1.2.1 (w/ or without mpd). It
> seems like it uses LOTS of all2all comms. Has someone tried it w/ Scyld ?

Scyld Beowulf isn't tied to MPICH especially. Our changes are actually
quite slight and can be easily applied to just about any other message
passing system.

Many parallel systems will work just fine if you tell them to use 'bpsh'
instead of 'rsh'. bpsh will use the copy of the binary installed on the
frontend.

Other systems require a bit more help. In particular, systems which expect
to put a daemon on each node will need help -- as the remote systems
probably don't have a copy of the target binary to execute.

In these cases (for example PVM), you can do a couple of things. Adopt a
new build target (Paul Springer's BEOLIN PVM arch target, for example) or
replace the fork()/exec("foo") pair in the daemon with something more like

nodenum=bproc_currnode(); bproc_rfork(-1); bproc_execmove(nodenum,"foo");

We don't recommend doing the latter, but it can be useful to get an
existing system up and running -- thereby giving you some time to think
about how to fix the software better.

Regards,
	Dan Ridge
	Scyld Computing Corporation