Intel is finally shipping the 64-bit Itanium

[Bob Drzyzgula]

Bari,

With all due respect (I've designed multilayer PCBs and have even read much
of Howard Johnson's book on signal integrity analysis, so this respect is in
fact considerable), I think that the problem that I (and I suspect, many
others on this list) have with your statements here is that you would appear
to be redefining the market space in order to achieve what you predict will
be a more optimal solution. IIRC, the *entire* point of Beowulf clusters, as
they were originally defined, was to build high-performance
parallel-processing clusters out of COTS, commodity parts. The idea was that
you could take surplus or otherwise underutilized systems and/or components
and assemble them into a new aggregate which would enable you to do
scientific calculations which were not otherwise feasible given your budget
realities. This idea has been refined somewhat over time, and the success of
the approach has brought new funding that will often, if not usually, allow
the purchase of the latest, greatest COTS hardware for new clusters, but I
think that for most of us the core philosophy still holds.

Here, on the other hand, you are talking about doing custom board designs
which are specifically optimized for cluster computing. This is a fine
approach to scientific computing, and you are of course not the first person
to do this, but I'm not sure that, in doing so, you fully appreciate the
extent to which you are departing from the mainstream of the Beowulf cluster
computing philosophy, as opposed to mere technology.

In the end, it may in fact be true that you will be able to build and sell
turn-key computing clusters which are competitive, from a price/performance
perspective, with turn-key clusters built from COTS motherboards and
chassis, and to do so in such a way that the bulk of the Beowulf software
base can still be used. However, part of the attraction of Beowulf clusters
to many organizations is the extent to which one can leverage available
staff and student labor, the creativity and frugality of determined but
under-funded scientists, below-the-line budget expenses such as power and
HVAC (never underestimate the value of getting another department to foot
much of the bill :-), as well as sunk costs such as data center space freed
up by shrinking or disappearing mainframes, to vastly reduce the apparent
acquisition cost for high-performance computers.

One major consideration is that, through the use of parts, the most
expensive of which may cost about $500 or so, many organizations (mine in
particular) will find that a cluster can be purchased out of operating
funds, rather than as a capital expense. This can give an extraordinary
boost to the organization's flexibility down the road. $100,000 spent on a
turn-key cluster might have to be capitalized -- and prove useful -- over a
period of three or four years. OTOH, spread that cost over a bunch of budget
line items, use student labor and racks from Home Depot or Costco, and you
might just be able to put the thing together without the bean counters even
noticing. Moreover, the cluster that was slapped together, if it proves not
to do the job or has outlived it's usefulness after a few months, can
probably be torn apart and the parts used for some other application. This
kind of insurance policy against fiscal error is extraordinarily difficult
to replicate for capital purchases of large, turn-key systems.

Although there are several high-profile and many lower-profile instances of
whole clusters being purchased and delivered on a single manifest, my
impression is that the vast majority of Beowulf clusters continue to be
built on an ad-hoc, piece-meal basis (am I wrong about this?). While I
applaud your creativity in searching for new ways to improve the
price/performance of available computing solutions, I think that it's pretty
important to keep in mind that the many people on this list who are building
the latter sort of cluster will remain quite unconvinced of your approach
until (a) they can build their own system nodes from your boards at an
*absolute* price competitive with those they are building today (note that,
for some of us, once the node cost crosses a threshold from operating and
into capital, they might as well cost a million dollars each, so
price/performance is only one part of the equation), (b) they have a
reasonable level of confidence that the parts they acquire from you will not
be so specialized as to be virtually worthless outside of the context of a
Beowulf, and (c) that your design will not be so unique that the
expandability and maintainability of their cluster would be shot to hell in
the event your company goes belly-up. Not that we don't wish you well...

Best regards,
--Bob