[rescue] SS10/20 death

[Francisco Javier Mesa-Martinez]

On Wed, 6 Feb 2002, Dave McGuire wrote:

> On February 6, dave at cca.org wrote:
> > >  Some mainframes use a "tiered" system for memory.  Erm, "storage".
> > 
> > Didn't one cray (Cray-2?) have two seperate memory systems? (And I'm
> > not talking about an SSD.)
> 
>   Not that I'm aware of, at least not one of the PVP machines, which are
> the ones I'm familiar with.  If one did, I don't think it could have
> been the Cray-2.
> 

That is correct only for the CRAY-1 (and maybe some of the later machines,
XMP/YMP?) had a single memory system. No memory hierarchies for these
puppies, i.e. no cache and no virtual memory. Well not 100% correct if we
think of the vector registers as a level of storage hierarchy, then you
have 2 memory levels in an original  Cray, the registers and the main
memory.

The idea behind the lack of memory hierarchies is that a Cray was not
meant to be a general purpose multiuser system. Besides having a cache and
v.mem adds inconsistency to the performance of certain instructions. And
that was a big no-no for Mr. Cray. Timing was very tight in the Cray-1
architectures, and having non-uniform memory accesses would have messed
things up. So cray just solved the problem by providing a huge memory
that was fast and had a huge bandwidth in order to feed the vector
registers (at least for that time).

However the Cray-2 had 2 levels of memory (3 if we include the vector
registers), each vector processor had a local memory (16KWords) which was
not accessible to the user unless you used some specific library calls and 
it can be thought of as some sort of cache, but I think you had to
specify the calls in your code in order to use such memory. All 4 processors 
shared a 512MWord memory (around 4GB which was very large back then, hell
even today it is pretty large... so imagine 15 years ago!). The cray-2 did
also lack VM system. Memory access was pretty shitty for the Cray-2
anyhow.. but it was relatively uniform.

I think the XMP and later introduced the SSD, which were solid state disk
storage units, but I do not think they were part of the directly
addressable memory (perhaps scratch or pseudo-swap state). But I have
never dealt with these machines...