[mpiwg-tools] PMPI and sessions init

William Gropp wgropp at illinois.edu
Tue Nov 10 09:27:01 CST 2020

Yes. We do not assume POSIX or Unix, and we try to stay away from specifics that can impact scalability (e.g., does this *semantically* require every process or node in a 10000000 node system to independently load files or save a GB of address space).  Some things do, and some do not.  But great care is needed to avoid tying MPI to particular implementations.


William Gropp
Director, NCSA
Thomas M. Siebel Chair in Computer Science
University of Illinois Urbana-Champaign

> On Nov 10, 2020, at 9:03 AM, Wesley Bland via mpiwg-tools <mpiwg-tools at lists.mpi-forum.org> wrote:
> We’ve explicitly avoided requiring specific mechanisms like dlopen so we can support the widest variety of tools on the widest variety of systems. dlopen, for instance, won’t work with static libraries, right?
> Thanks,
> Wes
>> On Nov 9, 2020, at 8:11 AM, Joachim Protze via mpiwg-tools <mpiwg-tools at lists.mpi-forum.org <mailto:mpiwg-tools at lists.mpi-forum.org>> wrote:
>> If dlopen causes libdl to load a library (i.e., the library is not yet
>> loaded and can be found and loaded), libdl calls the library constructor
>> as part of dlopen.
>> Forcing the MPI library to load (dlopen) all tool libraries before the
>> registration window closes, gives all dynamic tools the chance to call
>> the registration function, before MPI will continue initialization.
> _______________________________________________
> mpiwg-tools mailing list
> mpiwg-tools at lists.mpi-forum.org
> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-tools

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-tools/attachments/20201110/24107a64/attachment-0001.html>

More information about the mpiwg-tools mailing list