[Mpi3-rma] Draft Plenary Presentation for MPI RMA

Torsten Hoefler htor at illinois.edu
Fri Feb 4 12:35:23 CST 2011

Hi Jim,

> These look good.  Here's some feedback:
> Slide 4: RAW (Read after Write), WAR, WAW are commonly used terms for  
> this, should we use those?
This is my terminology. We can switch to this is nobody objects. I'll
also update the proposal.

Btw., the renaming from MPI_RMA_query to MPI_Win_query is still pending.

If I don't hear anything by *tonight*, I will update the proposal
tomorrow with the two changes above.

> Slide 21: UPC allows one-sided allocation of distributed shared data  
> through the one-sided call "shared void *upc_global_alloc(size_t  
> nblocks, size_t nbytes)".  I think Dan was pointing out that this is  
> impossible in MPI-2.  (Still don't think we can support this in MPI-3.)
You're right, allocation is either local or collective, never remote.

> Slide 22: UPC doesn't allow sharing data on the stack, only heap data  
> can be shared.  In general, this is kind of an iffy thing, might  
> encounter some resistance to suggest doing this.  Linux is going to  
> initially map the stack to the zero page which is marked copy-on-write,  
> so beyond the parts you have used, the stack doesn't exist.
That's what registration is for. It has to be created then. Or am I
missing something?

> Slide 30: Conflicting relaxed operations issued by a single source will  
> appear in program order when observed by that source.  IIRC, a third  
> party need not observe this same order.  I think that the result of  
> conflicting relaxed operations issued from different sources is 
> undefined.


 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