Microsoft Releases Computational Cluster Technical Preview To olkit

[Robert G. Brown]

On Mon, 5 Feb 2001, David Grant wrote:

> If I was a betting man, I bet they don't release the source code.....

...and thus we'll have an excellent opportunity to see if our long-held
belief that open source really matters.  Let's see.  We all know from
long (and sometimes bitter) experience that cluster performance is
related to things like:

  a) quality and tuning of the network stack
  b) quality and multitasking capabilities of the kernel, especially its
ability to handle high interrupt and context switch loads
  c) whether the particular versions of e.g. PVM or MPI employed are
"broken" in the region of operation of our application
  d) whether the parallel programming of the application itself is
correctly done or broken in some way (independent of the underlying
network stack, kernel, and parallel libraries used)
  e) whether the compiler and systems libraries used are in any way
broken with respect to their interface to any of the above
  f) arcana like CPU cache structure and speed, stride tuning (per node,
mind you), memory speed and management, bus speed(s), other bottlenecks
and nonlinearities

A weakness in any of these dimensions will lead to poor performance (or
no performance at all). As I understand the description, many of the
components of the Microsoft solution are proprietary and indeed owned by
diverse groups and integrated by Microsoft.  Microsoft is already justly
famous for its "it's not our problem" attitude to service.  Who's going
to be responsible for even IDENTIFYING failed components, let alone
fixing them?  A brief perusal of the beowulf list archives reveals that
problems are often nearly unique, and can cost far more to fix in human
time than Microsoft is likely to make even from a moderately expensive
package, especially if they have to pay second sources for service.

Perhaps this is intended to be a loss leader for Microsoft?

I personally am rather amused by all this.  I have on previous occasions
remarked upon what I'd like to call the "monopoly software development
cycle" (MSDC) -- Microsoft's delightful tendency to allow others to take
all the risks and do all the critical creative development work, prove a
valuable new market, whereupon Microsoft clones it from spec (using
their flotilla of high-quality development talent) and then move in and
market it as if they invented it.

Their next move (if they follow their usual pattern in the MSDC) will
be to fundamentally alter the paralllel libraries and support tools used
on their clusters (only) so that users that buy in originally and use
their libraries and tools will gradually find themselves unable to get
out without paying for an expensive port >>back to the open standard<<.
MPI was created in response to a demand for a common language across
parallel platforms.  We'll see how long it is before Microsoft uncreates
it in their own image.

Of course, in the past they have ALSO been able to do things in the MSDC
like undercut competitors' pricing by giving their "reinvented" software
away with the operating system for a year or so (as long as it takes for
the competitors who invented it to dry up and blow away).  This was
often until just after they released a major OS upgrade, as they could
ensure that their product actually worked after the upgrade while their
competitors' products historically were broken for at least a few
months, losing them customers and critical cash flow to pay for all that
original, high risk development effort.

In the fat days past they could afford to do this for years and (if
necessary) across multiple Windows upgrades.  As as I said, as long as
it took to grab oh, half the market or more.  They weren't too greedy.

How is the MSDC strategy going to work against linux?  At no time in
the future will they ever be able to undercut the price.  Linux
clustering solutions start at cost zero, and the most expensive ones
generally charge precisely a fair margin for value added (by e.g.
turnkey vendors).  Furthermore, linux turnkey vendors live in a highly
competitive market already -- they survive by providing direct,
personal, rapid, high quality service backed by considerable
consultative expertise.

Microsoft cannot put these guys out of business.  In fact, if Windows
clustering ever gets to the point where it can be taken seriously, they
can offer turnkey windows clusters to precisely the same customers that
are likely to buy them -- most of their added value is just as valuable
to these customers for a WinXX cluster, a linux cluster, a freebsd
cluster, or for that matter some of the dedicated iron clusters (e.g.
SP's).

Of course when they price out these solutions, guess which ones will be
the cheapest and why.  If the turnkey vendor is willing to stand behind
a linux solution and guarantee it on the same basis they guarantee a
WinXX solution, why waste money?

I therefore predict that Microsoft will lose money with this unless and
until somebody writes a killer app that runs only on Microsoft clusters
and that somehow isn't clonable to linux clusters.  Obviously, they're
the only ones who will invest the time needed to write such a thing,
since a) anybody else has nothing to lose by writing it to run on
generic MPI or PVM clusters including beowulfs; b) historically, if they
DO write a killer app for Windows, they will be shut out of their own
new market by Microsoft as fast as they can clone it and run through
the MSDC.  Besides, the open source development cycle is even faster
than the MSDC as it involves more contributors, smarter people, and the
memetic vigor produced by open interchange of code memes via the
internet.  Their killer app will be cloned in less than no time from
spec, and the OSDC will ensure that they only make money from it from
fools who LIKE to pay money to Microsoft.

I don't see them getting a lot of the scientific market -- Universities
love linux even for desktops even with site licenses that reduce per
seat costs to a few tens of dollars for MS products.  Linux management
scales nearly perfectly and it comes ready to run with a full suite of
incredibly useful apps that have to be pieced together at great cost on
MS boxen.  Government hates to waste money...

> Exxxcuuuse me, Mr. Gates, I need to license the additional nodes on my
> cluster, how much for the additional licenses?? :-)

...and then their is the cost scaling....

   rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu