[Beowulf] commercial clusters

[Vincent Diepeveen]

----- Original Message ----- 
From: "Angel Dimitrov" <stormlaboratory at yahoo.com>
To: <beowulf at beowulf.org>
Sent: Tuesday, September 26, 2006 8:10 PM
Subject: [Beowulf] commercial clusters


>
> Hello,
>
> I have some experience of running of numerical weather models on
> clusters.
>
> Is there many clients for processor time? As I saw the biggest
> supercomputers in the World are very busy! I'm wondering if it's
> worthwhile to setup a commercial cluster. Intel are planning for new
> processors - two CPUs each with quad cores. Two such machines will
> have power like one 50 GHz CPU:-)

Some years ago one could buy in system time at some swiss system with 
dolphin interconnects
(quite fast latency) for 50 euro a week a node (dual k7 nodes).

For those who really need a lot of crunching speed, they simply have to buy 
a system themselves.

Additional a professor isn't going to get a lot of money to buy in system 
time at a commercial offering.

Instead he can get free system time at national supercomputers.

There would be a commercial interest in clusters in a different way. In a 
practical manner.
In fact this has been proven already. That's for customers who do not want 
to run for months on your machine, but who just want to run for like a few 
hours at your machine at a specific time with a specific software 
application.

A good example is a chess cluster. However you HAVE to have then not only a 
highend cluster,
but also a program that works on your cluster which is a major problem. No 
need for many nodes,
jsut fast latency and the Seymour Cray principle.

The incentive for users then is that sometimes they are in urgent need for 
the most accurate analysis of a certain position. Then they simply want to 
toy with a program at home (so they need a gui at home that connects over 
the internet) and they just analyze as if it is a local program.

I'd be interested in setting up something like that. I do not have the 
hardware, but can deliver software for something like this. It will be good 
publicity anyway hitting newspapers and in case of a match
against a GM with such a cluster also hitting several TV stations. With 
statements like: "x times deeper searching than deep blue (completely true 
statements)".

Always good publicity as it makes you look intelligent and bright.

Of course such system time gets booked then for thousands of clients who 
just want it a couple of hours and first take a look whether at their time 
such a cluster is in usage.

The Seymour Cray principle however is very important in setting up such a 
cluster of say a node or 16 with each node a quadro core machine or a dual 
opteron dual core 3Ghz or dual woodcrest 3Ghz.

Or even better like 8 quad opteron dual core with a very low latency 
network.

That last is the fastest of course.

In fact count at 2 rails that you'll need. 1 for streaming from node to node
and 1 just for the latency.

One of the hardest things to parallellize on planet is search.

It will kick butt of course. People are prepared to pay for things that kick 
butt. In fact there is also sheikhs in the sport and some people with quite 
some money. A few of them would probably rent the machine for months at a 
time *after* it existed. They just want to toy and kick butt.

Giving people the feeling of kicking butt others, they are prepared to pay 
incredible amounts of money for.

But would you take the risk of setting up something like that?
I don't have the money for the hardware.

There is 1 billion people on planet earth who know the rules and know how to 
play chess.

In eastern europe it is one of the most popular sports. Every big company 
uses chess to let itself look like intelligent. But nearly none of all those 
companies sponsors the sport. They all sponsor soccer or tennis.

They just misuse it in that sense.
Use without paying.

In that sense the crunching number clusters will either be completely 
government dominated with a few exceptions of like oil companies who need 
some crunching speed. Of course the oil companies all go use basically 
gigabit ethernet. If you compare the crunching power the government has for 
its universities versus the total crunching speed in non-government hands, 
then there is a big big big difference.

So seemingly buying big sporthalls full of machines just to run 1 
application faster is in general not so extremely financial clever.

If i compare the average government cluster with the average companies 
cluster, the first big difference
is the price.

If a company DOES have a cluster, it's usually the utmost cheapest possible 
node with ethernet 100mbit (nowadays 1 gigabit is usually the maximum) 
connected with some real ugly slow router, versus
the cluster in government hands having a fast expensive network and each 
node of course an overpaid dual Xeon P4 (yes even in 2006 they use that 
outdated hardware).

Yes prices of 5000-6000 euro a node are very common.

In fact that's nearly more expensive than the new price of my dual opteron 
dual core with a raid10 array and with an ultrium2 streamer with a lot of 
tapes.

So just renting that government cluster against the price it has cost, is 
already a total nonsense idea in itself. It was in the first place way way 
too expensive when bought.

A small cluster with good network following the Seymour Cray principle is 
rather interesting for
certain purposes, big 1000+ node clusters is great for number crunching but 
unless you have a friend
in a commission of the government who can decide for you that you can 
deliver that cluster, there
is no money to be made from companies there.

On other hand if you are prepared to take risk in other way, feel free to 
email me.

Because alternatives like using a cluster to play quake (or whatever the 
latest 3d game is called)
faster, is going to give you porting problems.

*that* would sell too.

play for 25 euro an hour flight simulator. Bomb the evil Iranian regime 
(with plenty of targets in the 3 million square kilometer) in such manner 
that they cannot launch a single rocket at the aircraft carriers!

Now in full 3d bla bla bla.

The problem is the videocard and the internet connection then.

The only realtime things that is possible to do remote is things like chess.



> Any ideas and comments are welcome!
>
> Angel
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
>