Re: Plea for simplicity

Eric Salo (salo@mrjones.engr.sgi.com)
Wed, 1 Nov 1995 16:40:53 -0800

> Similar to SUNMOS, Puma is designed for high performance
> computing on the compute nodes of the Paragon. It
> is performance oriented and supports only basic Unix functionality.
> Demand paging is not supported. However, it does support
> timesharing on the compute nodes, including the necessary
> protections. For instance, a process can create a "window",
> but only processes in the same application are allowed
> to write to this window, unless another application is given
> explicit permission to use the window. Do we need to
> specify some protections for MPI get/put?

So if you were running a.out and b.out in some master/slave application, would
the a.out process be able to write into the memory of the b.out process? I
don't think that any additional protections are needed for MPI beyond
communicators.

> Under Puma, an application can shoot itself
> in the foot, but it is not allowed to harm the OS or
> other applications. Puma does not protect an application
> from itself.

Sounds good.

> Does your SHMALLOC allow me to specify the 'real' address?
> Or will it simply malloc space out of my heap and return
> this 'real' address?

It will come off of the heap. Locally, the address returned by MPI_SHMALLOC()
is a real pointer to a new memory region. When this address is passed to
MPI_PUT() and MPI_GET(), magic happens to translate the local address into the
appropriate remote address so that the bits can be sent to the correct place.

-- 
Eric Salo         Silicon Graphics Inc.             "Do you know what the
(415)390-2998     2011 N. Shoreline Blvd, 7L-802     last Xon said, just
salo@sgi.com      Mountain View, CA   94043-1389     before he died?"