On Nov 25, 2005, at 5:31 PM, Salman Pervez wrote:
>>> The tcp rpi module accesses certain global variables
>>
>> ... which is fine as they are global per process and there is no
>> interference with other accesses.
>>
> thanks for the response. This makes sense except for one thing. The
> request list, referenced by lam_rq_top, is this local to each procecss
> also? Or is it a global list accessible by all processes? In case it is
> local to each process, at what point are requests from peer processes
> transferred to the appropriate request list?
It is global in each MPI process (hence, no locking).
Remember that LAM/MPI is layered software -- the LAM layer is the
run-time environment, and it has no knowledge of MPI. So the lamd's
are purely LAM processes. The MPI layer uses the services from the LAM
layer and implements the MPI standard. Hence, the share/mpi directory
goes into libmpi (as well as all the RPI modules and several other
things), and is therefore specific to MPI processes.
--
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/
|
