I am interviewing Michael K
Patterson, PhD, PE, Senior Power/Thermal Architect, Eco-Technology Program
Office, Digital Enterprise Group at Intel.
(That’s quite a title - I am glad I was able write that rather than say
it outloud…) Michael:
1. I
understand you have written a white paper about the energy efficiency of thin
client desktops. Can you briefly talk
about that? (when it was published, where it can be found, etc.)
Yes, Lorie Wigle and I
co-authored a paper for the Electronics Goes Green 2008 conference in Berlin last
September. The paper, titled
“Alternative Compute Models: End-to-end energy impact”, looked at the total
energy impact, including the servers and extra data center energy required for
thin clients for a number of different compute models. So we looked at thin clients, desktops, and
laptops.
2. Your research concluded
that thin client desktops are not energy efficient when considered end to end,
right? It sounds counter-intuitive so
perhaps you can help me out with this.
Well, no, that’s not exactly what we
found. The results showed that thin
clients, laptops, and desktops all can be very energy efficient when done
correctly. The key finding was that thin
clients are not always the most efficient model, even though there has been a
lot of press claiming they are. In fact
the most efficient model was a laptop.
Of course to keep it fair, if you add an LCD display to the laptop for
in-office use things are a lot closer.
For desktops (and the laptop) the key is using
the latest CPUs with the best efficient performance, LCDs instead of CRTs, and
actively managing the clients and turning on power management features.
The thin client model energy efficiency
depends on a couple of factors. First,
how many thin clients can be supported per server? Second, how efficient is the data center that
those servers reside in? Let’s focus
there a minute. A typical data center
Power Usage Effectiveness or PUE is 2.0.
That means that the total power going to the data center is 2.0X times
how much goes to the servers. That extra
power goes to cooling, power distribution, and lighting. All these energy uses need to be attributed
to the thin client model. As I started
with, the other key factor is how many clients per server. Our analysis showed that in an inefficient
data center (PUE=3.0) you needed about 60 clients per server for the thin
client model to be as efficient as the desktop.
In a very efficient data center (PUE=1.5) you only needed about 30
clients per server to make the thin client model equivalent. Interestingly, when compared to a laptop,
even with an LCD, you would need to run over 100 clients per server to have the
same efficiency.
And that’s the main issue for the thin client
model. It all depends on the work being
done by the employee. If it’s just data
entry or terminal-like work, it can be fine.
But as soon as you ask the server to support 30 clients all running
typical office applications like word processing or spreadsheets it really
slows down. Then if you get into any
more intensive applications using graphics, streaming media, or advanced
analysis, you really need to have far fewer clients per server. It’s all in the paper and the references, but
at a very high level a rule-of-thumb would be that for today’s knowledge
worker, doing more than just data entry, a core-per-client is a good
target. So a quad core, dual processor
server could support 8 clients with roughly the same level of performance as
the managed desktop. At these levels,
thin clients are the least efficient model.
3. Were you surprised by your own conclusions?
In any of this we should not
judge what’s best until we do the whole analysis. And of course, energy efficiency is only part
of the analysis. TCO, security, and
manageability all come into play, as does the ability to work remotely. When I presented the paper in Berlin I joked that I
wish I had a thin client. Everyone
looked at me puzzled until I explained if I had a thin client I could have
watched the movie on the flight over instead of catching up on my email.
Each site needs to do their own detailed
analysis. But when you think of how
precious the data center space is, these results shouldn’t surprise anyone. I have heard people who have to build and pay
for them say “let’s get the Powerpoint out of the datacenter”. When you do the math, it makes a lot of
sense.
4. Can you give specifics on better solutions for
call centers, doctor’s offices, and other traditional thin client
environments?
For these typical thin client
applications, thin clients can be a great choice. Again one needs to look at the whole
picture. In any situation where thin
clients are being considered, my best advice is “pilot”. Set up a pilot, at the proposed number of
clients per server, and see how the employees fare. If there is no negative
impact the thin client model should get a very strong look. On the other-hand, if the thin client stalls,
or can’t keep up with the dedicated desktop or laptop you need to be very
cautious. The productivity loss will far
outweigh any energy efficiency or other TCO gains, particularly when summed
over the life of the thin client installation.
5. What would be the best way
to obtain energy efficiency gains throughout an entire IT life cycle, i.e.,
where the data center is just one aspect of the entire picture?
When one talks to someone
about the possibility of a thin client application, the first question to be
asked is “What’s the PUE of your data center?”
If they don’t know, there are a whole lot of areas for efficiency gains
before they should even consider the thin client model. The big challenge is not rushing into
something that just sounds good. And one
of the problems here is that the impact, once it’s been implemented, may never
be measured. The impact to the site
energy can be very difficult to determine.
The data center energy use and the office space energy use are not
usually measured together or compared, the impact may be mixed with other
changes in the building or IT inventory, data center loading, or even the
seasonal variations on the heating or cooling loads. So it’s really best to do the detailed
analysis up front. The best way to gain
energy efficiency in the IT space is to use the latest hardware, get rid of
CRTs, enable the power savings features, and manage it all centrally. That’s true for thin clients, desktops, and
laptops.
6. Well of course that doesn’t include embedded
energy, and I know Intel is studying the impact and that recycling more
computer hardware is uppermost in your research. Separately, do you have ideas on added
original research to be done on this topic?
Hmmm, original research? I can’t think of any specific scientific
research needed on this as much as we need better tools to help the end-user to
do the full analysis. If all we do is
look at the specific watts of an individual desktop, laptop, and thin client
and pick the thin client because it is lowest, we could very well regret that
choice. That’s where more work can be
done, tools for analysis.
Thanks so much, Michael! Great to talk to you…
and likewise Deborah, it’s always a treat to
talk to you. You always push me; with good questions and though provoking
ideas, thanks!
HAVE A GREEN DAY!
Corrections: It was me (Jonathan Koomey) not Bob Metcalf who showed the graph about data center electricity use being comparable to Mexico. I did NOT say that growth in data center electricity use was speeding up (in fact, it's slowing down as the industry focuses on the issue). And I also did not say that sending a letter uses 300,000 times more energy than emailing a PDF--I compared the mass of material in the electrons needed to send the PDF to the mass of a 5 g sheet of paper--that's where the factor of 300,000 comes from. It's an illustration of the power of dematerialization, but it doesn't translate directly to energy.