I apologize for dragging in this conversation in a different
direction, but I'd be very interested to know why the behavior with
the Playstation is different from other architectures. The PS3 box has
a single gigabit ethernet and no exapansion ports, so I'd assume it's
behavior would be no different than, e.g. a regular PC using the TCP
BTL. Perhaps it has something to do with the Cell BE architecture,
then. What was the reasoning behind this decision?
I am keen to know about such 'hybrid' parallel programming paradigm,
e.g. using Cell BE or NUMA or CUDA on top of an MPI (or even a grid
topology). I'd appreciate any pointers to any material in this
On Sun, Nov 15, 2009 at 4:48 PM, George Bosilca <bosilca_at_[hidden]> wrote:
> By default only one socket per peer per physical network is opened. However,
> Open MPI has the possibility to open multiple socket per peer per network,
> based on some experiments with the Playstation (where having multiple socket
> allow for more bandwidth). The MCA parameter that allows such behavior is
> On Nov 13, 2009, at 17:59 , Charles Salvia wrote:
>> When using TCP, how many sockets does each process open per peer-process?
>> Does each process open a single socket to connect to each peer-process, or
>> does it use TWO sockets, one for sending, one for receiving?
>> -Charles Salvia
>> users mailing list
> users mailing list