[Beowulf] MS HPC... Oh dear...

Mike Davis jmdavis1 at vcu.edu
Wed Jun 14 09:50:54 PDT 2006


The HPC market is a far cry from the game console market. Having served 
my time in both publishing and Marketing (in past lives) I can 
understand your point.

But, I don't believe that the point is completely relevant. For example, 
even if Sun offered me their e2900 for the same price as a quad dualcore 
opteron, it wouldn't make sense for my cluster implementation. Though 
its certainly a better machine for certain applications (for example 
assembling a genome of multiple chromosones each wth 250k 1 kilobase 
sequences), it wouldn't be as good at Gaussian, or Gamess, or vasp, or 
Monte Carlo.

I own e2900's and I own opteron, OS X, and Xeon clusters. They each have 
their strengths and weaknesses. Apples may be easier to set up for 
life-sciences, but they are a bear for vasp. The xeons tend to be 2/3 
the speed of the opteron even though they run at faster clock speeds. 
The e2900's are stuffed with 64 and 96GB of ram, making them great for 
apps that can't be efficiently broken down to run on the 2-8GB of RAM 
available per cluster node.

Finally, while marketing does play a role, we have seen a major surge in 
cluster useage without a huge marketing arm being involved. We buy 
clusters to solve problems. If I thought that a Windows cluster was the 
way to solve a given problem I would consider it. However after being in 
computing for 20 years, I know that the key to marketing lies in what is 
left unsaid. MS will not talk about patch Tuesday in their marketing. 
Somehow the idea of losing 24-36 hours on a cluster because of a bad 
patch strikes me as a dangerous and expensive proposition. 4800 to 7200 
processor hours is ALOT of work.

Mike Davis

Vincent Diepeveen wrote:
> ----- Original Message ----- From: "Joe Landman" 
> <landman at scalableinformatics.com>
> To: "Thomas H Dr Pierce" <TPierce at rohmhaas.com>
> Cc: <beowulf at beowulf.org>; "Robert G. Brown" <rgb at phy.duke.edu>; 
> <beowulf-bounces at beowulf.org>
> Sent: Wednesday, June 14, 2006 2:02 PM
> Subject: Re: [Beowulf] MS HPC... Oh dear...
>> Hi Thomas:
>> Thomas H Dr Pierce wrote:
>> [...]
>>> Microsoft HPC will work IF the market wants the technology that they can
>>> supply.  The answer is in market segments. And the Supercomputer market
>> I don't think the technology is what will win people over.  Its the
>> packaging.
> The only thing that sells products very well is Marketing.
> Mass marketing of course only is possible when the products are easy to 
> use for companies,
> and not overexpensive.
> When users feel they get ripped off, then you can of course only try to 
> sell such a plan to a government organisation,
> but no longer to companies.
> Let's compare mass marketing in highend.
> Even though i'm not sure you want to call that highend.
> I do :)
> xbox360 from microsoft has as a problem that there is good alternatives.
> Take for example Sony who will probably be real good in selling their 
> playstation3, as they are a marketing giant.
> Bigger marketing giant than microsoft actually.
> This where xbox360 sells for 100-200 cheaper than ps3.
> Important is the customer price in the end. The PS3 is affordable. 
> Cheaper in fact than a full blown computer
> The graphics card is seemingly quite a bit better of ps3, it has in 
> total 9 cores versus xbox360 has 3.
> PS3 has a massive amount of software that runs at it when compared to 
> xbox360.
> Porting from PS2 to PS3 is relative easy. Porting from xbox to xbox360 
> will be a nightmare for the programmers.
> So there is plenty of reasons to assume that ps3 will get a major success.
> Now let's move to highend.
> Microsoft is completely unchallenged there.
> There is not a single other giant marketing themselves onto TV.
> Only TV campaigns work to really mass market products or get your 
> company known.
> I can remember microsoft marketing xbox. I can remember sony marketing 
> all kind of products of them onto TV.
> However not many highend 'giants' can compete here with microsoft.
> Let's list m$ advantages:
>  - about every program runs on their platform
>  - they are cheaper than anyone else (about every product is online at 
> p2p networks) and
>     legal licenses they are cheap too
>  - it's easy for programmers to develop software for the platform
>  - massive marketing advantages there is for windows server as it gets 
> marketed as 1 big
>    product that serves all companies.
> The only disadvantages i can think of are of pure technical and short 
> term nature, not financial nor practical.
> Such as: "they don't run yet at a 1024 processor system".
> Well let me assure you. Be happy they don't yet. If they WOULD, then you 
> can pack your bags already.
> The price of a ferrari F1 car is about 5 million dollar a piece or so. 
> That's just raw material of it.
> Once they made a 'road' version of it. Only a few of them were sold, at 
> a real expensive price.
> You just can't compare that with the tens of millions of other cars the 
> so called  'cluster of the masses',
> those all run windows.
> Soon it'll have a good name when the next edition gets released.
> Then it'll have soon better drivers than linux too for highend network 
> cards,
> who no longer will release.
> And in the end what matters is that visual studio 2005 and newer is 
> simply 20% to 50% faster for most codes
> than GCC. GCC suffering from bugs everywhere with pgo still and somehow 
> it isn't really faster for K8.
> Who will be so stupid to port his parallel program to MPI?
> Waste of time! They work for windows or they don't work at all!
> Let's all grow up. MPI will be soon history when m$ takes over.
> If Myri signs some exclusive contract with m$, perhaps m$ will implement 
> some 'compatible' form of messaging into the windows
> kernel, which takes care that windows parallel software for clusters in 
> future could get ported perhaps to MPI with some big work.
> In all other cases it won't get ported to MPI at all.
> Now you can of course still keep running your years 80 software and hope 
> it'll keep working in future, but in how far all this
> is realistic is not real visible. One day you'll retire and then there 
> is not many left that will run that software.
> [...]
> Vincent
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf

More information about the Beowulf mailing list