Hi Tomasz,
On 6 May 2017 at 18:46, Tomasz Buchert wrote:
| On 03/05/17 17:54, Ralf Stubner wrote:
| > [...]
|
| Hi all,
| this is really another iteration of #741297.
|
|
| Honestly, I believe that the whole test of openmpi "existence" using
| dlopen is unnecessary. This code is only executed if we have compiled
| against openmpi, so why do we have to double-check that it is in the
| system? I propose to completely drop the dlopen test.
While true for us, it is not always true for Rmpi on other system so upstream
for Rmpi added this. I haven't heard from him a while.
But I do recall that we needed this for some other braindeadness with the
multiple shared library -- I distinctly remember fighting this for many
months. The problem, really, is the need for RTLD_GLOBAL | RTLD_LAZY in
order follow symbols into the other libraries which OpenMPI decides to split
this over.
dlopen("libmpi.so.1", RTLD_GLOBAL | RTLD_LAZY)
So we would need to demonstrate that removing this block does in fact work.
| I attach a patch that does exactly that. I tried to prepare NMU, but
| had a hard time with version 1.0 of this package :(.
What version 1.0? There never was one for Rmpi.
Dirk
--
http://dirk.eddelbuettel.com | @eddelbuettel | [email protected]
