[Mpi3-rma] Conflicting accesses

Torsten Hoefler htor at illinois.edu
Sat Dec 11 18:50:55 CST 2010


Hi all,

We discussed the issue with the lockall/shared mode that the outcome of
overlapping puts or put/get will be undefined. Our discussion ended at a
point where we believed that such accesses are not valid in
lockall/shared because there is only one access/exposure epoch.

The statement (not in MPI-2) at the end of page 42 seems to imply that.
However, I don't think that anything on page 10 (the rules for
conflicting accesses) mandate this. Jim created this interesting example
and both of us think it should be legal to do (without unlocking as was
required in MPI-2):

lock(exclusive)                                                                 
get(tail)                                                                       
flush()                                                                         
put(tail)                                                                       
unlock()    

It should be true because flush() implies completion at the target which
means that the "mini-epoch" is now finished (yes, in MPI-2, we needed to
close an epoch to get completion, however, in MPI-3 we don't).

Do we all agree? I'd like to remove the sentence starting with "For
example" on page 42.

All the Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler         | Performance Modeling and Simulation Lead
Blue Waters Directorate | University of Illinois (UIUC)
1205 W Clark Street     | Urbana, IL, 61801
NCSA Building           | +01 (217) 244-7736



More information about the mpiwg-rma mailing list