Page 265 -

P
236
236 P a r t V : a r t V : T T h e G r e e n i n g P r o c e s sh e G r e e n i n g P r o c e s s

750 MHz server 750 MHz server 750 MHz server 750 MHz server

One server’s power can be
subdivided into smaller,
individual units

750 MHz server 750 MHz server 750 MHz server 750 GHz server

Four-core,
12GHz server
750 MHz server 750 MHz server 750 MHz server 750 GHz server

750 MHz server 750 MHz server 750 MHz server 750 MHz server
FIGURE 12-2 For this example, a four-core server delivering 12GHz of power can be divided into sixteen
750MHz servers.

But it gets even better. If half of those servers are in off-peak or idle mode, the remaining
servers have that much processing power available to them. So, in essence, if half of the
servers are barely being used, the remaining eight servers have 1.5GHz available to them.
You can’t expect the servers to run at optimal usage all the time. That’s because the
more the servers are called on to deliver this capacity, the less responsive the system will be.
Rather, a good rule of thumb is never to let a server exceed approximately 50-percent
sustained utilization, thereby reserving capacity for moments where usage needs may peak.
For instance, let’s say that a group of servers are somewhere between 1 and 5 percent
utilized and that they run different applications that are used at varying levels during
different times of the day, week, year, or season. If you were to virtualize those servers onto
a single machine so that you have about 50 percent utilization, you should be able to use
peak capacity for the various applications during their individual peak times. All it takes is
a little bit of planning to save a lot of energy and reduce waste.
Storage is another issue with virtual servers. Although we tend to look at CPU usage as
our yardstick, input/output (I/O) considerations still need to be added to the equation. I/O
overhead for storage and networking can gum up the works. If your server has high storage
or hardware I/O requirements, it might be better to leave it on its own physical server.