[MPICH] Mixing MPI

David Hart dhart at indiana.edu
Wed Dec 21 11:31:17 CST 2005


This reminds me of the PACX-MPI project.


On Dec 21, 2005, at 10:05 AM, John Robinson wrote:

> Our application forks new processes off a normal TCP socket  
> connnect/fork listener process, which then MPI_Init and connect to  
> a cluster over MPI using connect/accept calls.  Multiple of these  
> frontend processes talk to the cluster at the same time (multi- 
> threaded processes).  So we have in effect a client/server model  
> within the cluster.  The connect/merge semamtics are an adequate  
> model of a TCP listen/fork model for cluster communication.  We  
> have the client as process 0 after the MPI merge and it sequences  
> the cluster protocol with broadcast commands.
>
> I found it quite easy to get this working.
>
> /jr
> ---
> David Minor wrote:
>> What I did is create a server that communicates with it's clients  
>> using
>> sockets. It is also the MPI rank 0 process. It has a separate thread
>> that handles the sockets from the threads that handle the MPI calls.
>> This is over simplistic but MPI wasn't really designed for this  
>> kind of
>> work (unfortunately). David
>> -----Original Message-----
>> From: owner-mpich-discuss at mcs.anl.gov
>> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of andy
>> Sent: Monday, December 19, 2005 12:32 PM
>> To: mpich-discuss at mcs.anl.gov
>> Subject: [MPICH] Mixing MPI
>> I am writing a server program using MPI and other packages which use
>> sockets and pipes and the like. I have had a ponder and a google but
>> cannot see how to introduce MPI to the server using something like  
>> the
>> unix select() to avoid polling. Can I get hold of file  
>> descriptors? Or
>> am I simply looking at what must be a fairly common task in the wrong
>> way? Thanks for any suggestions.
>




More information about the mpich-discuss mailing list