[Mpi-comments] One comment on MPI-3.0 Draft 2, August 2012

George Bosilca bosilca at eecs.utk.edu
Wed Aug 15 15:27:27 CDT 2012


In other terms it depends on what you prefer to understand by fairness.

geeks: http://en.wikipedia.org/wiki/Unbounded_nondeterminism

otherwise: http://dictionary.reference.com/browse/fairness

  george.

On Aug 15, 2012, at 21:29 , William Gropp wrote:

> This is a good point.  Using the Testsome or Waitsome interface permits an application to respond fairly without burdening the implementation with any particular fairness definition.  Not every code can be conveniently written this way, but it addresses many needs.
> 
> As has been noted, specifying the exact meaning of fairness (e.g., round robin) may have serious performance impact, as an implementation may make fair progress but in a different, possibly non-deterministic, order.  Even if "fairness" is provided, it should be defined in terms of the essential properties, not a particular implementation that happens to have those properties.
> 
> Bill
> 
> William Gropp
> Director, Parallel Computing Institute
> Deputy Director for Research
> Institute for Advanced Computing Applications and Technologies
> Paul and Cynthia Saylor Professor of Computer Science
> University of Illinois Urbana-Champaign
> 
> 
> 
> On Aug 15, 2012, at 8:54 AM, Jim Dinan wrote:
> 
>> Hi Sebastien,
>> 
>> Something to consider in addition to probe is that fairness semantics are also relevant to test/wait operations that take arrays of requests.  It may make the most sense to attach a fairness attribute to communication contexts (communicators and windows) to provide more flexibility.  This could already be done by an MPI implementation through an info key on a communicator/window, however implementations need not support all info keys and fairness could be a requirement for applications that rely on it.
>> 
>> Cheers,
>> ~Jim.
>> 
>> On 8/15/12 10:19 AM, Sébastien Boisvert wrote:
>>> Maybe there could be something like MPI_Init_fairness which could either
>>> be MPI_FAIRNESS_NO (the current behavior) or MPI_FAIRNESS_YES (which
>>> would do round-robin reception).
>> _______________________________________________
>> mpi-comments mailing list
>> mpi-comments at lists.mpi-forum.org
>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-comments
> 
> 
> _______________________________________________
> mpi-comments mailing list
> mpi-comments at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-comments





More information about the mpi-comments mailing list