[mpich-discuss] Behavioral change for MPI_Ibarrier(MPI_COMM_SELF, &r) in mpich-3.2

[William Gropp]

I agree.  This is a perverse interpretation.  While I understand the optimization, the nonblocking routines should return an active request, not a null request.  (The real optimization is probably in not creating the request, rather than freeing it; while requests should be managed in such a way that creation is cheap, its likely that defensive coding will initialize the request, which can be non-trivial.  Still, this doesn’t seem worth the surprise to the user.)

Bill

William Gropp
Director, Parallel Computing Institute
Thomas M. Siebel Chair in Computer Science
Chief Scientist, NCSA
University of Illinois Urbana-Champaign





On Dec 7, 2015, at 9:29 AM, Jeff Squyres (jsquyres) <jsquyres at cisco.com> wrote:

> I understand the optimization here, but I think Jed does have a point: it is weird.
> 
> Further, MPI-3.1 5.1.2 p196:38-39 states:
> 
>    "A nonblocking call initiates a collective operation, which must be completed in a separate completion call."
> 
> Arguably, this does not directly state that you need to return a valid request from any of the NBCs.  But it's a reasonable interpretation to believe that it should.
> 
> Further, Jed is right that this might be a micro-optimization (i.e., avoiding the creation or deallocation of a request), but it doesn't help much unless the app also does an associated "if" to test and see if the request==MPI_REQUEST_NULL and therefore doesn't call TEST*/WAIT* (i.e., that would be a further micro-optimization -- which probably isn't negated by the app's "if" because TEST*/WAIT* have to do the same "if" for REQUEST_NULL).
> 
> In short: yes, this is a micro-optimization.  But:
> 
> a) it kinda falls in the "violates the law of least astonishment" category
> b) is deallocating a free listed request that expensive?
> c) if you have to explain this every time, are you really optimizing anything?
> d) it's a single process case -- it's already damn fast. Do you really need to optimize it further?  :-)
> 
> (To be clear: I don't really care; I just bring up these points to highlight the other side of the story)
> 
> 
>> On Dec 7, 2015, at 6:39 AM, Jeff Hammond <jeff.science at gmail.com> wrote:
>> 
>> Completed requests are set to MPI_REQUEST_NULL.  Since MPI_Ibarrier (and MPI_Barrier) is a no-op for a communicator of size 1, it is trivially complete and it would seem that MPICH is optimizing away the unnecessary request object here.
>> 
>> Why don't you use MPI_Test instead of direct comparison?
>> 
>> Jeff
>> 
>> On Sun, Dec 6, 2015 at 10:04 PM, Jed Brown <jed at jedbrown.org> wrote:
>> In MPICH 3.2, MPI_Ibarrier(MPI_COMM_SELF,&r) returns r ==
>> MPI_REQUEST_NULL.  This is also the case for other communicators of size
>> 1.  Previous versions of MPICH and also all versions of Open MPI return
>> a valid request not equal to MPI_REQUEST_NULL.  This bit me because it
>> was natural to compare to MPI_REQUEST_NULL in a conditional for an
>> Ibarrier/Issend/Iprobe algorithm.  It can be worked around with a
>> slightly more complicate conditional, but is this change intentional?
>> 
>>  The call returns a request handle, which must be passed to a
>>  completion call.
>>  -- MPI-3 §5.12
>> 
>> This does not say "unless the communicator has size 1".
>> 
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
>> 
>> 
>> 
>> -- 
>> Jeff Hammond
>> jeff.science at gmail.com
>> http://jeffhammond.github.io/
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
> 
> 
> -- 
> Jeff Squyres
> jsquyres at cisco.com
> For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
> 
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss