Supercomputing in Small Spaces
As the super-computing community continues to aggregate larger and
larger clusters of machines, the cost and size of the resulting
supercomputer increases tremendously.
Where as the traditional "price / performance" ratio addresses the
question "What is the highest performance supercomputer I can afford
to buy?" it does not address the more important question "What is
the highest performance supercomputer that I can afford to own?"
The cost of owning a supercomputer can be broken down in the
following way. Note that the traditional price / performance metric
only includes acquisition which is one of the many costs associated
with owning a supercomputer.
-
One-time costs
- Acquisition
- Facility construction
-
Recurring costs
- Facility maintenance
- Power for operation
- Cooling
- System administration
- Failed hardware
- Security costs
Because of the magnitude of the costs beyond those necessary for
acquisition, especially for top of the line supercomputers, we feel
that the total cost of ownership / performance is the best metric
for comparing various supercomputing options.
At SC 2001, we demonstrated the ability to build and run real codes
on a 15-GFlop 24-node Transmeta 3U form factor cluster made by RLX
Technologies. Although the peak performance was less than what would
have been obtained by using higher performance CPUs, the cooler
running Transmeta chips allowed higher density, along with lower
cooling and power requirements. Thus dramatically lowering the most
tangible of the total costs of ownership.
What was not demonstrated was the ease with which the cluster can be
installed and administered. All the nodes are hot swappable and
required no tools for installation. The administration software
makes configuration fast, flexible and easy from a remote location.
Finally, the reliability the cluster has be exemplary.
