[mpich-discuss] MPI_Win_create synch

[Alberto Giannetti]

It doesn't. I was just wondering how processes are returned the  
handler to the correct window region when multiple memory windows are  
created in the master. For example:

if( rank == 0 ) { // Master, create memory
   ...
   Mem = MPI::Alloc_mem(memSize, MPI_INFO_NULL);
   winQueue = MPI::Win::Create(pMem, memSize, sizeof(int),  
MPI_INFO_NULL, MPI::COMM_WORLD);

   Mem2 = MPI::Alloc_mem(memSize2, MPI_INFO_NULL);
   winQueue2 = MPI::Win::Create(pMem2, memSize2, sizeof(int),  
MPI_INFO_NULL, MPI::COMM_WORLD);
   ...
} else { // Slave, get handler to remote memory
   ...
   winQueue = MPI::Win::Create((void*)0, 0, sizeof(int),  
MPI_INFO_NULL, MPI::COMM_WORLD);

   winQueue2 = MPI::Win::Create((void*)0, 0, sizeof(int),  
MPI_INFO_NULL, MPI::COMM_WORLD);
   ...
}

How does the MPICH implementation guarantees that the slave first  
call to Create returns an handle (winQueue) to pMem and the second  
call (winQueue2) to pMem2? Without using global identifiers or  
synchronization, the two slave Create calls could be assigned in  
reverse order: the first call could be assigned to the remote memory  
region pointed by pMem2, rather than pMem.

I don't see any id to match the two window segments in the master and  
the slave, so I think there is a barrier implicit after each Create()  
collective call?

On Jul 7, 2008, at 11:28 PM, Rajeev Thakur wrote:

> Win_create is a collective operation. All processes in the  
> communicator
> (COMM_WORLD in this case) must call it. Does it hang if you call a  
> second
> Win_create in both master and slave?
>
> Rajeev
>
>> -----Original Message-----
>> From: owner-mpich-discuss at mcs.anl.gov
>> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Robert Kubrick
>> Sent: Monday, July 07, 2008 9:23 PM
>> To: mpich-discuss at mcs.anl.gov
>> Subject: [mpich-discuss] MPI_Win_create synch
>>
>> I want to create a window region in my master process. Slave
>> processes are connecting to the region for some one-sided
>> operations and I don't need to create extra windows in each slave.
>> The code will look something like this:
>>
>> if( rank == 0 ) { // Master
>>    ...
>>    pMem = MPI::Alloc_mem(memSize, MPI_INFO_NULL);
>>    winQueue = MPI::Win::Create(pMem, memSize, sizeof(int),
>> MPI_INFO_NULL, MPI::COMM_WORLD);
>>    ...
>> } else { // Slaves
>>    ...
>>    winQueue = MPI::Win::Create((void*)0, 0, sizeof(int),
>> MPI_INFO_NULL, MPI::COMM_WORLD);
>>    ...
>> }
>>
>> How does MPI match the correct window handle in the slave.
>> Lets say I create more than one window in the master, which
>> window handle is returned to the slave after Create()?
>> I noticed that if I add Create() calls to the master, the
>> program hangs, I think because it must match the extra
>> windows in each slave.
>> Is there a sort of barrier embedded in each Win::Create()?
>>
>> Rob
>>
>>
>