On Tue, 21 Nov 2017, Dirk Eddelbuettel wrote:
On 21 November 2017 at 12:06, Allin Cottrell wrote:
| On Tue, 21 Nov 2017, Riccardo (Jack) Lucchetti wrote:
| > On Tue, 21 Nov 2017, Dirk Eddelbuettel wrote:
| >> I _could_ change the standard build to always use openmpi. May open yet
| >> another can of worms, but at least it is less involved.
| > Would gretl_x11 work on a system without openmpi installed, or would it
| > complain about the missing library?
| It will work OK, as will gretlcli. It's just gretlmpi that won't work
| without the MPI library.
| We've taken some pains to arrange this: neither gretl_x11 nor gretlcli
| are linked against the MPI library; we get the required symbols at
| runtime via dlopen, if they're available.
That sounds good.
So ... one never needs one of the MPI frontend, ie orte_run? Or, conversely,
one gets this only when gretl is launched this way? Ie user opts in, no side
effects for others?
Yes, it's an opt-in thing. We need a back-end that supports "mpirun"
or "mpiexec" -- e.g. orte_run -- but only if/when gretlmpi gets
called, not for the regular programs.
Can't stand in the way of such Pareto-improvements ;-)
Heh, that's right.