[Mpi3-rma] RMA examples

Torsten Hoefler htor at illinois.edu
Thu Nov 11 22:02:00 CST 2010


On Thu, Nov 11, 2010 at 08:55:46PM -0600, Pavan Balaji wrote:
> [...]
> Correct: Most folks
> Not correct: No one
> Abstain: Bill
>
> Torsten was somewhere between abstain and correct.
Thanks for posting Pavan!

Just to clarify, I was abstain at the beginning, but after we seemed to
accept that some (ncc) architectures with strong progression cannot be
supported in the RMA unified model, I think those should be ok.

This makes the lock-all-shared mode and the proposed lockfree mode
semantically equivalent. And as before, I find it non-intuitive to start
a shared lockall epoch to perform (essentially unlocked) global random
accesses. 

Also, it seems like we need to allow conflicting accesses in such an
epoch (which would be undefined). To allow this, we would need to change
the definition of shared lock, which right now reads: "Updates to the
window are protected by exclusive locks if they may conflict.
Nonconflicting accesses (such as read-only accesses or accumulate
accesses) are protected by shared locks, both for local accesses and for
RMA accesses.". Changing this would obviously be possible but seems to
conflict with the literature on shared/exclusive locks, which suggests
to use shared locks for non-conflicting accesses and upgrade to
exclusive locks if conflicts might occur. MPI-2.2 is in line with this
definition if we relax "lock" to "transaction" and assume that the
granularity of a "lock" is a whole window.

All the Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Whenever you put data on a computer you lose some control over it. And
when you put it on the Internet you lose a lot of control over it.
(Schneier about facebook.com, 2006-09-21).



More information about the mpiwg-rma mailing list