[Info-vax] The Future of Server Hardware?

[Stephen Hoffman]

On 2012-10-01 08:31:25 +0000, JF Mezei said:

> Woopty doo, so their cooling costs will be lower.

That's part of being more efficient.

In a classic data center, you'd roughly figure what you put in for 
power would need to come out as heat; that you'd size cooling by the 
full-powered compute load.

Cooling efficiency here is improving, and substantially.

> But they still have
> tens of thousands of servers, a number large enough they still need to
> locate near a very large power generation facility.

So?  That much power with as many mainframes would want to be close to 
cheap power, too.
Because the folks that build these data centers are not idiots.

> All these "green" data centre stories focus on reducing the cooling
> costs. They never discuss the actual server power costs and of course
> never compare having tens of thousands of servers versus having X
> mainframes with comparable compute/IO power.

So look around.  Why aren't mainframes everywhere?   Maybe because 
they're not cost-effective for these loads?

> Perhaps the millenium generation has never heard of mainframe computers
> that can run linux and really think the only way to scale things is to
> just had thousnads more PCs in your data centre.

Mainframes require massive power, and are less efficient when throttled 
and operating at partial loads, and are very cost-inefficient at 
dealing with load spikes.

Mainframes are also more expensive due to the redundancies that are 
inherent in a one-egg-one-basket server design.

You still get the same scaling issues and distributed computing 
considerations once you exceed the load a single mainframe can support, 
too.

Smaller compute units are also in-rack disposable, where you'd need to 
keep staff around to repair the bigger servers.

Smaller servers are also more immediately scalable for longer-term 
increases in load, and can be installed by the shipping container as 
load requirements increase.  And can be replaced by the shipping 
container, as the servers age out in terms of reliability and power 
efficiency.


> You'll note that established mature companies who are not flush with
> cash like the millenium companies such as Google, Facebook, Apple,
> Twitter etc) do not run out and build data centres with 10,000 PCs

Different markets and different applications have different computing 
requirements.  (Duh.)

> If there were a truly objective "green" study comparing power costs to
> run mainframes against equivaent number of 1U servers (equicalent in
> computer/IO power), what would it show ?

It would show that different business can and should pick the 
appropriate technologies for their requirements, and make the most 
efficient use of them.

Do you seriously think that senior Google management doesn't look for 
ways to lower their power bills?

Put another way, the folks building these data centers and these 
applications are not idiots.




-- 
Pure Personal Opinion | HoffmanLabs LLC