[Foundation-l] Fundraising season launch

Neil Harris usenet at tonal.clara.co.uk
Tue Oct 10 23:48:02 UTC 2006

David Gerard wrote:
> On 10/10/06, Brad Patrick <bradp.wmf at gmail.com> wrote:
>> If we have +/- 250 servers running now
>> *and we double in traffic every four months
>> *and the costs per server (for round number purposes, including racks,
>> power, etc.) are $4000
>> *we are talking about *starting* with a $1M computer order being
>> fulfilled before March 1.
>> If we add to that the next 500 computers, at the same rate, we are
>> talking about a batch of $2M beyond that.  /me's head spins.
> Heh. What's the upfront and service contract on an IBM minicomputer -
> make that two clustered in different cities - running as many copies
> of Linux/390 as we need?
> (I wonder if IBM would sponsor that, just to say they had.)
> - d.

IBM don't sell their mainframes as low-cost compute farms, and that's 
because mainframes simply aren't _that_ powerful relative to their 
price; what they are useful for is consolidating lots and lots of 
relatively lightly-loaded services into one _highly_ reliable and 
maintainable system. That reliability and maintainability is what you 
pay for when you buy a mainframe; and if you really need it, you need to 
get a mainframe. (There are, for instance, stories of mainframes being 
rebuilt like George Washington's axe until none of the original machine 
was left, whilst still in service, without stopping any programs from 
running or taking any filesystems down: they were simply shifted from 
processor to processor, and disk to disk, with the system still running.)

For example, a fully-loaded 54-CPU top-of-the range IBM z9-109 
"enterprise-class" mainframe is claimed to be able to do "a billion 
transactions per day". Now, if we assume that each of those transactions 
is equivalent to serving a Wikipedia hit, that's about 11000 hits per 
second, about 60% of the load of 18000 hits per second currently served 
by the several hundred CPUs in the Wikipedia cluster. However, if the 
price list below is anything to go by, it'll set you back about $22M for 
the privilege, which is considerably more than the entire Wikimedia 
Foundation budget to date.



However, if you wanted a high-price/performance-ratio compute farm from 
IBM, I'm sure they could oblige in other ways: for example, they 
currently hold the first, second, and third places in the TOP500 
supercomputer listing with their closely-coupled cluster systems, so I 
would imagine they could probably quote you for a web server farm.

You could always ask them for a quote: as a high-prestige project, 
there's always a possibility they might be inclined to do a good deal on 
it for the PR benefit... but I'm willing to bet they'll try to sell you 
a cluster system rather than a mainframe.

-- Neil

More information about the foundation-l mailing list