Skip to main content

Green S+S Dynamic Live IT ‐ Stop Climate Change Now???!!! 

Posted By:  Giovanni Marchetti 
Publish Date: 7/17/2008

“Green IT” is a catch‐all name for a series of initiatives, all aiming ultimately to save money with
the excuse of (or by means of) reducing energy consumption by datacenters, hence their
“carbon footprint”. I can think of several questions to understand how “green” we are. As for
the answers, I need time to find them.

1. What and where do you measure?
OEMs typically expose power consumption data via their management cards (where they are
available), but those data are limited to the server in question. As its workload varies, its
consumption will vary. Also, there is a “minimum” power consumption needed to keep the
server running. Although this has decreased significantly in modern machines with ACPI, power
stepping circuitry, it is still a pretty high percentage of the theoretical maximum. Switches,
racks and “environmental” systems (lights, a/c) also consume power, but few of them actually
have a way to tell how much. So, where do we measure and how do we aggregate and analyze
the measurements?

2. Have you got the appropriate hardware for the job?
For instance, blades have become very popular because of their reduced footprint. Most of
their power is allocated to the CPUs (obviously, as their is little else on the blade). For CPUintensive
applications, that is not a problem. If you also want to run data‐intensive applications,
you will likely have to store those data elsewhere, maybe on some sort of SAN. Storage systems
typically have no way of knowing what kind of application uses the data and when, hence they
keep their disks spinning all the time. From a power utilization point of view, in this case it is
more efficient to keep the data local to the system, hence a blade may not be the best solution.

3. Is your software making optimal use of such hardware?
Your workload can be optimally distributed on the hardware to minimize power consumption
and make best use of what is available. Let us suppose that you have a memory‐intensive
application on a NUMA (e.g. Opteron) machine. For best performance, you will try and allocate
1 thread of such application per CPU socket, thus optimizing the use of available memory
bandwidth. This leaves the other cores on those sockets idle. On Windows today you cannot
turn them off; active power management (processor power states, if enabled in the bios, are
supported by server 2008) will slow them down.

4. Further reading
This list of questions is definitely not exhaustive. As I look for answers, more will probably come
to mind and I’ll make a new “feature of the week” out of it. In the meanwhile, feel free to send
me your comments.
Here are a few pointers that may help inform a discussion:
- Lewis Curtis’s blog - Little Miss Enviro-Geek - The Green Datacenter Blog- Microsoft’s Environment web page- IBM Green Datacenter paper- Green Computing Paper- The Green Grid