Reading Gregory Pfister's In Search of Clusters and then talking to
Wolfpack participants intrigued me with the person who brought together and
explained the ideas necessary to implement clustering. I wondered whether he was
aware of the impact of his book and what he thought about how vendors where
putting his concepts into practice. The following interview resulted from my
interest in this book and its author.
Q: Are you aware that the name Wolfpack is
based on your book?
Yes, Microsoft sent me a T-shirt and a letter letting me know
that Clusters influenced the name.
Q: When you subtitled your book, The
Coming Battle in Lowly Parallel Computing, what did you have in mind?
I was thinking of clusters of commodity PCs vs. large SMP systems
(12-, 16-, 32-way). The battle lines are just now being drawn.
Q: Do you see a large market even for
Wolfpack Phase 1? Is it possible to see a tenfold increase in the quantity of
clusters shipped in the next year?
It's possible, but many vendors will have to participate before
clusters reach critical mass. Can one large vendor accelerate the process?
Absolutely.
Q: How does Intel's four-way Pentium Pro
take advantage of NT's SMP capabilities?
Going to the Pentium Pro, Intel made several changes that
significantly increase the bus capabilities, which is the key limiting factor in
SMP. So SMPs made with the Intel bus and the Pentium Pro chipset will be a lot
better than SMPs made from Pentium chip set. They will be formidable machines.
Q: Will four-way clusters provide a better
price/performance point than large (8, 16, 32 CPUs) SMP boxes?
Large SMPs cost more to produce than clusters of small SMP boxes
and will therefore be more expensive. However, until cluster vendors provide a
single-system image of the cluster, a large SMP will be easier to administer
than a cluster.
Q: As we move toward scaling clusters,
will message passing between nodes become more important?
The Message Passing Interface (MPI) is an existing standard that
would help this effort, but I imagine Microsoft will adopt its own proprietary
solution, Distributed Component Object Model (DCOM).
Q: What are the challenges that face
Wolfpack as it moves toward scaling clusters?
The single-system image will be a huge problem to overcome. Will
NT support a clustered file system to let a user view a set of files that are
the sum of all nodes in the cluster? Will communication from outside the cluster
have to know the address of an individual node, or will a message be able to
point to a cluster? Will network management systems see the cluster as a single
element or as multiple separate computers? A whole lot of these services will
have to be available before an administrator will say that a cluster appears as
a single computing resource in the same way a large parallel system appears. The
company that provides that stuff first and on the best price/performance
hardware will be a winner.
Q: Is everything coming together now that
will let clustering become commonplace?
Possibly. There's no new technology necessary to do this. The
technology exists. What's necessary is the will and the vision to do it.