[Beowulf] fast interconnects, HT 3.0 ...

[Christian Bell]

On Wed, 24 May 2006, Richard Walsh wrote:

> Greg Lindahl wrote:
> >On Wed, May 24, 2006 at 09:17:58PM +0200, Eugen Leitl wrote:
> >
> >  
> >>pGAS assumes shared memory.

Global addressibility in pGAS cannot assume shared memory, or else
pGAS would not be available for clusters.

> >>    
> >
> >There's not much point having this discussion when we disagree on
> >such elementary terms.
> >
> >-- greg
> >  
>    Perhaps a reminder that:
> 
>           global addressability does not entail global cache coherency
> 
>    but, we all new that anyway, right?

As for global cache coherency, we agree. The X1 is a "native" pGAS
machine and implements global addressibility *without* enforcing
global cache coherency -- the "native global address space" only
caches accesses to *local* memory, remote reads/writes are not
cached.  Global cache coherency runs into some of the scalability
limitations that the Altix faces.

cheers,

    . . christian