[Beowulf] Electricity cost: a critical survival issue of our ICT infrastructures.

Mark Hahn hahn at mcmaster.ca
Mon Apr 8 17:46:19 PDT 2013


> Energy cost will massively grow within this and next decade because of
> decreasing oil and gas supply and rapidly growing electricity demand.

I'm not sure where you get this.  in north america, for instance,
there has been a fairly dramat increase in available energy (from new
sources of natural gas, mainly), while electrical demand has been
*falling* for years (at least in my part of Canada).

> For the time from 2008 until 2030 the electricity comsumption growth
> of all our ICT infrastructures has been predicted by a factor of 30

well, on the one hand, that's a 1.2% compounded growth rate,
which seems reasonably low.  on the other hand, only a complete ass
would believe anyone can predict such numbers >20 years in the future.

than again, I'm at least as uncertain about what "ICT infrastructure" means.

> --- if current trends continue. Also the price per kWh will probably grow
> by another factor,

I hope so.  energy is far too cheap to discourage its use, and our 
environment suffers badly as a consequence.

> so that the electricity cost to run our ICT 
> infrastructures
> may grow by a factor of 100 or even more during this period of time.

I assume you mean "by 100%" rather than "by a factor of 100x" - 
this would imply that the cost of power more than triples.
that would surprise me a lot, given that power prices (wholesale,
at least) have decreased over the past decade.

it's also true that ICT has gotten dramatically more efficient
in the past few years - sometimes in absolute numbers, but certainly
in constant-workload comparisons.  I'm not really talking about 
escascale here - that's a prestige thing, driven by political, 
not engineering concerns (I'm pretty sceptical about imputed scientific 
motives, too.)

> We need your ideas on how to cope with this challenge by current trends.

I think the industry is doing well on its own, though higher power prices
would provide more encouragement.  currently, it costs roughly 10% of the 
capital cost of a server to run it at 100% for a year (here, at least).
most sane IT environments plan for 3-ish years of service, so that's 30%,
and counting a mediocre PUE of 1.3, so let's call the total power component
a doubling of cost.  that's certainly big enough to affect planning processes,
even accelerated obsolescence.  (suppose, for instance, that even a generic
server, 2 years later, performs 4x faster at the same power and price.)

in short, yes there are new mega-datacenters out there, but they are 
causing a lot of consolidation - elimination of crappy old PUE 3.0 
datacenters with racks of big/fat 4U servers and missing blanking panels
in favor of decently packed multi-tenant VM hosts.  that trend is basically
irrelevant for HPC, our clusters are usually already airflow-optimized, 
fully-utilized, etc.

regards, mark hahn.


More information about the Beowulf mailing list