[gmx-developers] Re: gmx-developers Digest, Vol 99, Issue 14

[Peter Kasson]

Hmm--you've managed to confuse me (and I worked on some of the
parallelization schemes mentioned here).  So maybe more clarity is in
order.
A couple issues:  one is clarity is options, a second is what we want
to have as the supported and advertised options.  I agree with Michael
that the tMPI implementation is pretty darn good for most single-node
setups.  It just works, and it gives linear scaling in most cases
(interesting NUMA cases aside).  My impression was that we're not
quite at that point with OpenMP right now, correct?

We want to think about clarity, backwards compatibility, and future
compatibility.  Ultimately we want to conceptualize a certain number
of processes and a certain number of workers per process, right?
(Sort of a virtual equivalent of #nodes, ppn in some PBS headers.)

I think the general framework Berk suggests is good, but the
nomenclature is a bit confusing.  Perhaps the best way to do this is
keep np = np, have nt roll over to np if we're using tMPI (this is the
expected backwards behavior), and have something clearly distinct like
-nw = -nworkers for the number of workers (threads) per process? -If
there's an -nt and a -nth, that gets confusing.  -ntsm presupposes an
implementation.

Cheers,
--Peter