Hi Jed, The MPI standard doesn't mandate a particular behavior in case of an error, so doing what the user asked for is also allowed. :) In seriousness, I completely agree on the importance of error checking. The coverage in RMA has improved significantly through the MPI-3 development effort, and we are continuing to work on it. ~Jim. On 2/8/13 1:16 PM, Jed Brown wrote:
On Fri, Feb 8, 2013 at 1:08 PM, Jim Dinan <[email protected] <mailto:[email protected]>> wrote:
However, it will work fine in MPICH. In MPICH 3.0, I redesigned the synchronization error detection; as of 3.0, this is the only invalid RMA synchronization where we don't flag an error. This is not so much because we choose to support this as an extension to RMA, but because the flush state needs to change collectively, making this difficult (not O(1)) to detect.
If this is really the only instance, it may not be worth it, but in general, it's helpful to have a run-time (or even compile-time, if necessary) option that does the strictest checking you're willing to implement, even when it implies a performance hit. Certainly useful for people trying to determine whether they have interpreted the standard correctly, or when given someone else's code that is failing mysteriously.
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss