[Mpi3-rma] RMA proposal 1 update
Torsten Hoefler
htor at illinois.edu
Wed May 26 14:21:00 CDT 2010
On Wed, May 26, 2010 at 01:13:05PM -0500, Rajeev Thakur wrote:
> > > Not just lockall. The user would have to assert that no other
> process
> > > will call regular lock with an exclusive lock.
> >
> > I believe this would just be da-facto illegal in the current set of
> > rules. Because lock-all gives you a shared lock. If anybody
> > is trying to acquire an exclusive lock, he would deadlock.
>
> Not necessarily. They can just get serialized. For example
>
> Rank 0 Rank 9
> ------- ------
> Lockall(shared) Lock(rank 5, excl)
> Put Put
> Unlockall Unlock
>
> Either rank 0's operations happen first or rank 9's. If rank 0 never
> calls unlockall, it is a different story. But that can happen even today
> if you do a lock but don't call unlock.
Sure, I was talking about at the same rank (when we allow multiple
locks)
Rank x:
Lockall() // shared by default
Put
Lock(exclusive) // deadlock
...
It's clear that lockall is not collective (would be identical to
win-fence then).
All the Best,
Torsten
--
bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler | Research Associate
Blue Waters Directorate | University of Illinois
1205 W Clark Street | Urbana, IL, 61801
NCSA Building | +01 (217) 244-7736
More information about the mpiwg-rma
mailing list