Re: SMP scalability: 8 -> 32 CPUs

William J. Earl (wje@fir.engr.sgi.com)
Thu, 3 Dec 1998 10:03:51 -0800


Stephen C. Tweedie writes:
> Hi,
>
> On Mon, 30 Nov 1998 23:49:24 -0800, "William J. Earl"
> <wje@fir.engr.sgi.com> said:
>
> > The basic hardware components, however, have been shipping for some
> > time, so the "mechanism for binding 1024 cpus into a single box" does
> > exist.
>
> You can put a cluster into one box. You can bind the memory tightly
> through message passing, and you can present a single system image to
> the user. The fact that it's in one box doesn't make it SMP!

The many 128P systems we have shipped and the two 256P systems
being tested are logically SMP systems with a single kernel image, not
a cellular cluster, and are not partitioned. The extra memory latency
for the router network is generally smaller than local memory latency,
so the worst-case disparity in memory latency is less than 2 to 1. A
128P system is no more or less a cluster than a 4P Origin 200 or an 8P
Origin 2000. They all have more than one memory controller, with 1 or
2 processors per controller, and access to remote memory is via the
network. Memory latency is not as good as on a 1P desktop system, but
it degrades very little on average with increasing processor count,
and the aggregate bandwidth scales up with processor count. The
Origin is certainly not an SMP at the hardware level, and the memory
latency is somewhat non-uniform, but, for most applications, one can
use it as if it were an SMP.

It is possible to reconfigure a 128P as a cluster (in several boxes,
though, not just one, since a 128P is a big machine) using partitioning,
but that is not how most of those systems are used. The "SSI over a
cellular cluster" version of IRIX is not presently available, so any
partitioned Origin systems are really just ordinary clusters, except for
lower inter-partition latency.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/