Mailinglist Archive: opensuse (3506 mails)

[James Knott <james.knott@xxxxxxxxxx>]

M Harris wrote:
> On Wednesday 20 September 2006 06:12, James Knott wrote:
>> As a for-instance, Big Idea company... you
>>
>>> know, the Vegietales people, made their Jonah movie by networking 500
>>> linux machines together (all rack mounted pcs) rendering day and night
>>> for a couple of months... they didn't buy a mainframe with 500 LPARs....
>>>    :)
>> At a Linux conference I attended last April, an IBM rep made the point
>> that for a given load, mainframes tend to be more energy efficient than
>> a bunch of PCs, an important consideration these days.  And since the
>> presentation was about virtual machines, He said that it's easier to
>> balance loads across several virtual machines, than separate boxes and
>> that "networking" between virtual machines runs at memory speed, which
>> is considerably faster than ethernet.
>      More energy efficient???    Is he kidding?  Have you ever seen the power 
> supply rooms for a mainframe shop...?  Have you ever seen the water coolers 
> and air-conditioners needed to keep those silly things from melting? 
>      There are several factors to consider:
> 
>   1) The mainframe is expensive... we're not talking just spendy here... we're 
> talking outrageous.
>   2) Each lpar operates at a fraction of the speed/efficiency of a stand-alone 
> counterpart.  (not energy, computational power)
>   3) The mainframe is housed in a single location -duh- and if you loose the 
> power (or anything else) you've lost everything. (all eggs in one basket is a 
> bad plan...  ancient prairie proverb)
>       [ On the other hand: ]

Clusters tend to be in one location too.
> 
>   1) Distributed systems can be built with off-the-shelf parts 
> inexpensively... many science centers have done this already... 
>   2) Distributed systems do not require air-conditioners, water coolers, 
> etc... and can be placed strategically on the power grid so that loosing one 
> (power, etc)  doesn't bring the whole thing down.

Clusters tend to be in one location.  Otherwise communications will kill
you.  Even with individual computers, heat is always a problem, and
these days is severely straining AC.  Water cooling can reduce problems.

>   3) Distributed systems have cross redundancy as well as shared resource... 
> including processor resource... again, so that if you loose one machine the 
> others just take over and everything keeps running.
>   4) Distributed systems (say all 1.8Ghz) running in parallel (say 500 of 
> them) will blow the doors off of a mainframe running lpars... even using 
> semetric multi-processing. (or some other stat pulled out of the air)
> 
> The old school of thought was to take a mainframe and share it across multiple 
> users (terminals) and there-by defray the cost of the mainframe... because 
> nobody had personal computers!!  uh... 1977.
> 
> The new school of thought is to take thousands of networked super PCs (200GB 
> dasd +, 1.8Ghz +, ) and pull them together into one gigantic networked super 
> processor with total redundancy including processor resource.
> 
> The brain (if you will) is the network. Each cell or cluster of cells function 
> as one small redundant piece of the whole... and the whole collective body is 
> thousands of times more powerful than the sum of its many parts... 
> 
> The power of a mainframe is limited (divided) across lpars... 
> 
> The power of a distributed mega cluster is *amplified* on the whole.
> 
> ------
> 
> I have experimented with this at home... most folks have one large more or 
> less state-of-the-art pc with the latest toys attached... it goes down... and 
> they're in the shop for a while.
> 
> My system is a distributed cluster of 14 older systems that never goes down. 
> It is completely redundant, lightning fast, with self backup, mirroring, and 
> processor sharing.  And the cool thing is that authorized users (member of 
> the household) can logon from one of five terminals (also nodes in the 
> network) and access the entire cluster (or individual machines) including 
> graphical interfaces to any of the nodes on the cluster if need be. Its fun, 
> its fast, its reliable, and its cheap....
> 
> Now... consider multiplying this idea by 500 or 5000 or 5,000,000.  The Linux 
> copies running on an IBM mainframe is so 1970.s...  common guys... this is 
> the 21st Century... think of the Seti at home project... and you're coming 
> closer to where I'm going here.
> 
> There is a reason why Seti at home is not called Seti in a mainframe...   ;-)

IBM sells mainframes and IBM sells clusters.  I suspect they know the
pros and cons of each.  It's more a matter of which is best for your
application.  In some cases mainframe, in others clusters.  Clusters
tend to do best, when you can have many parallel operations, mainframes,
when you can't.