[Beowulf] SGI to offer Windows on clusters

[Robert G. Brown]

On Thu, 18 Jan 2007, Richard Walsh wrote:

> Ashley Pittman wrote:
>> On Wed, 2007-01-17 at 08:50 +0100, Mikael Fredriksson wrote
>>> Yes, it is.  And more so if this cluster/LAN can also utilize som type
>>> of "MOSIX" system.  This will substatially increase the throughput of
>>> "standard serial" processes.
>>> 
>> 
>> I find this statement hard to comprehend, how can any OS substantially
>> improve throughput of jobs unless what it replaces is incredibly
>> deficient in some way?  The limiting factor on clusters is the speed of
>> the hardware, even if some OS magically manages to be say 50% more
>> efficient doing it's bit than another OS it's still only a tiny percent
>> of time used, substantial improvements in job throughput can only come
>> about from better parallel algorithms, better code or faster hardware.
>>
>>
>    While I agree with this argument, especially at small scale, at very 
> large scale operating
>    system derived load imbalance (so-called skew, due to the random nature 
> of system
>    call driven interrupts) can destroy scalability, and thus efficiency. 
> This is worth mentioning,
>    although I would not expect Windows to improve on Linux in this context. 
> You need
>    a light-weight kernel like Catamount to reduce skew.
>
>    There is a very good paper showing the effects of skew at scale by 
> Kerberyson, et al from
>    Sandia.

It also isn't the point.  Nobody, and I mean nobody in this universe,
analyzes and compares WinXX and Linux from a performance point of view.
WinXX isn't engineered for performance per se, it is engineered for its
marketability first and foremost, its real or perceived
user-friendliness at the desktop (as a major component of its
marketability), its ability to permit less-than-genius admins to manage
a complex, networked, operating environment (marketability), and its
real or perceived availability of corporate support for same.
Performance is a distant last place, way down there beneath stability
(and look how well they do with stability) and ensuring that their
perception as the only platform that makes sense for the development of
commercial software remains unchallenged.

Stability is important to sales, especially in a server environment and
as I previously noted pro-grade systems people can make a WinXX Server
setup sufficiently stable for production purposes.  The (correct)
perception by developers that it is corporate suicide to develop for a
presumed corporate linux network that extends to the desktop is also
critical to MS's business plan and continued success.  Performance is a
competitive metric, and would only matter if they had any competition or
if their performance isn't "adequate".

Linux remains around the 1% level in US desktop occupancy, and even in
the pacific rim where its numbers are the best it only makes it to 3% or
so (the rest are doubtless mostly bootleg WinXX).  The largest monopoly
ever to exist in the history of the world laughs at these numbers.  On
the broader server market Linux fares better, but it is still very much
David against Goliath where David may appear sometimes to be winning,
but Goliath has yet to be hit in the head with any kind of stone.

At this point, with massive investment in MS stock on the part of
corporate retirement and pension accounts, hitting MS in the head with a
killer stone would probably trigger a nationwide panic or even a
depression.  It could still happen, but I really expect that killing
Goliath may take years of nibbling at his heels and not a single blow,
with Goliath fighting back and changing form all the way to try to avoid
his fate.

Basically, MS's cluster product is almost certainly designed to do two
things.  One is provide them with a credible presence in the cluster
market not because it is particularly important to them as a profit
center but because hurting linux and the other unices strengthens their
position in the general server market in many ways.  They do not want
corporate admins or execs to be forced into installing a linux cluster
in a WinXX server environment as it is way too likely that they'll learn
that Linux can take over their WinXX server responsibilities and save
them a bundle, once they've already invested in linux admins and gotten
over the startup "hump".  The other is to defend against the possibility
that a "killer app" might emerge in that marketplace, or that the apps
in that marketplace might attract enough software companies into the at
least part time linux development market that they, too, might start
looking at porting their software in general to linux.

>From what I can see, there are various things gradually lowering the
barrier between linux development and Windows development -- making it
easier to port Windows code directly to linux with a recompile, making
it easier to run Windows code directly within linux without a Windows
license.  Wine/cedega, vmware, win4lin and others on the one hand,
cross-architecture development libraries on the other hand.  Microsoft
has a strong interest in maintaining those barriers and doubtless moves
things around to keep code porting difficult (compare how easy it is to
move code between unices to how difficult it is to move between linux
and MS, and how much expense that adds to the task of maintaining a code
base in both worlds).

So the MS Cluster will doubtless run shrink-wrapped software in
corporate WinXX-only server rooms, permitting them to pay MS $500 a box
(or whatever) rather than hire another two or three $80K/year employees
to run a linux cluster, and control the cluster, policy, and the
application from the familiar Windows toplevel GUI.  For a single task,
and a cluster with only 16 or so nodes, that's a total bargain.  Even
linux cluster consulting companies would be hard-pressed to provide a
turnkey cluster and remote manage it (necessary to avoid hiring those
local linux admins) for only $8-10K in margin.  For larger numbers of
nodes, of course, at some point this argument doesn't scale, but even
there you come up against MS's formidable sales force, who will whack
you on the head with all sorts of TCO FUD, support FUD, ease of use FUD,
and ultimately may well convince you that it is easier and cheaper to
just pay them a 20-40% "tax" per node out to hundreds of nodes than it
is to take that shiftless, irresponsible, hippie-supported idealistic
"movement" that is linux into their pristine server rooms.

They'll win some of these arguments, lose others.  And not care much
either way.  Their purpose is served.  What they continue to risk is the
gradual movement of BIG companies away from Windows-dominated server
rooms and the even gradualler movement of a very few of those companies
towards a corporate linux desktop.  The last thing in the world they
want is companies publically trumpeting major savings from converting to
linux on an enterprise level.  Yet the barrier to doing such a
conversion is gradually coming down, and the number of big companies
with a significant linux presence grows.

So it isn't about performance, it is about presence.  SGI is a big, high
visibility deal and provides them with FUD-war ammunition, a way to slow
the bleed of server sites to linux, and which will almost certainly pay
for itself even if it is little more than a batch queue program with a
nice GUI and perhaps a STABLE API -- something linux could certainly
use.

    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