[Fwd: Re: [Beowulf] Cell in HPC]

[Greg Lindahl]

On Thu, Jun 01, 2006 at 09:07:03AM -0400, Mark Hahn wrote:

> > How do I effectively control placement when I don't know how big my
> > cache is?
> 
> I guess you mean you don't know because you can't count on getting 
> any particular fraction of the shared cache, right?

Yes.

> so just do these
> streaming/blocked reads directly into your per-core cache.  no?

We prefetch into L1 or L2, depending on what's going on. In this case
L1 is per core and L2 is shared, so you're suggesting we give up
prefetching/blocking in L2?

-- g