Global Shared Memory and SCI/Dolphin

James Cownie jcownie at etnus.com
Fri Jul 18 04:41:43 EDT 2003


> > In fact I thought all four of SCALI/Quadrics/Myrinet/Infiniband could do
> > RDMA ??
> 
> There's a terminology problem here: Some people mean cache-coherent
> shared memory, like that on an SGI Origin.
> 
> Another term for non-cache-coherent but globally addressable and
> accessible memory is SALC: Shared address, local consistency.
> 
> And yes, all 4 of Scali/Quadrics/Myrinet/Infiniband support the
> non-cache-coherent kind of shared memory. Programming models in this
> area are:
> 
>   * UPC: Unified Parallel C
>   * CoArray Fortran
>   * MPI-2 one-sided operations
>   * Global Arrays from PNL
>   * The Cray SHMEM library

However there's another axis to the classification which you haven't
mentioned, and which is also extremeley important, which is whether
the remote access is "punned" onto a normal load/store instruction, or
requires a different explicit operation.

I like to refer to the Quadrics' model as "explicit remote store
access", since it requires special accesses to (process mapped) device
registers to  cause remote operations to happen; therefore the process
making a remote access has to know that that's what it wants to do. It
can't just follow a chain of pointers and end up doing remote accesses
transparently.

Note, also, that AFAIK the explicit remote store accesses in the
Quadrics' implementation are cache coherent at both ends, so they are
not SALC. (Both because there isn't a shared address space, and
because they are consistent at both ends !).

As I understand it the Quadrics' model is that there are multiple
processes each with their own address space, but that by explicit
operations a process can read or write data in a cache coherent
fashion and without co-operation from its owner in any of the address
spaces. (At least that's how it worked back at Meiko ;-)

I suppose you could view the {process-id, address} tuple as a shared
address space, but it seems a bit of a stretch to me.

-- Jim 

James Cownie	<jcownie at etnus.com>
Etnus, LLC.     +44 117 9071438
http://www.etnus.com





_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list