[mpiwg-rma] sync on vector of windows?

Jeff Hammond jeff.science at gmail.com
Sun Oct 5 21:44:25 CDT 2014


Lock/unlock are problematic. We'd need trylock semantics to avoid deadlock. And the end result is something harder to use. 

And I don't think these ops fall on the critical path. PGAS-style usage is going to lockall once per window and unlockall just prior to freeing it. 

Jeff

Sent from my iPhone

> On Oct 5, 2014, at 7:09 PM, Min Si <msi at il.is.s.u-tokyo.ac.jp> wrote:
> 
> Looks good for me. How about lock{all} / unlock{all} ? An optimization could be combining internal lock/unlock messages especially in unlock{all} .
> 
> Min
> 
> 
> (14-10-5 20:57), Jeff Hammond wrote:
>> I created https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/459 in
>> response to my long-standing issue with this in ARMCI-MPI and the
>> additional issues encountered by Min (copied) in another project.
>> 
>> Please let me know what you think about this.
>> 
>> Thanks,
>> 
>> Jeff
> 



More information about the mpiwg-rma mailing list