Who runs large capability jobs
pauln at psc.edu
Fri Jul 28 12:06:37 PDT 2000
If this is the same SGI machine that i'm thinking of then it's
not a terrible machine in terms of scalability. But they made
some interesting design decisions based on older cray technology
which weren't so good. The machine arch vaguely resembles the T3D
in that there are 2 cpus per network connection into the torus instead
of 1 connection per cpu like in the T3E. The link bandwidth is about
half that of the T3E and serves 2 cpus. This network on the sgi is also
used for maintaining cache coherency, which can be very expensize on
the network. It's obvious that since most of Cray's intellectual
property was gobbled up by SGI that they have the tools and know how
to build a killer machine. But trying to keep costs down while building
a "one size fits all machine" isn't going to yield the best hpc machine.
On Fri, 28 Jul 2000, Victor Ortega wrote:
> On Fri, 28 Jul 2000, Greg Lindahl wrote:
> > I made a specific statement about scaling to 1,000 processors. A single SGI
> > machine does not scale that big; the biggest is 512, and the O3000 doesn't
> > raise that. The numbers speak for themselves. I don't know why we've had
> > this long thread about it. I hope someone got something useful out of it,
> > but it's been mostly people talking past each other.
> That the SGI machine scales to 512 sounds to me very much like a
> scalable system. Scalable, at least to me and perhaps to many others,
> means that the system can grow in power and throughput at a reasonable
> cost, and I don't think SGI is stretching the term or being misleading
> by calling their system scalable.
> If you had only said "Note that the SGI won't scale beyond 512", you
> would have provided useful information and no one would have argued
> with you; instead, you decided to rant on SGI for using the term
> scalable. You say yourself that the numbers speak for themselves; so
> why did you delve into the subjective topic of what scalable means?
> Please show a little humility in accepting responsibility for starting
> this long, acidic thread, and please try to be less offensive on this
> list. I appreciate the expertise you bring to this list, but I
> believe I am not alone when I say that the offensive statements and
> the hair-splitting aren't welcome.
> Beowulf mailing list
> Beowulf at beowulf.org
More information about the Beowulf