The peak performance characteristics of the CTC Velocity Cluster are impressive, and do raise high expectations. For example:

• Each Velocity v1 node consists of 4 500 MHz processors, capable of generating up to 4 billion floating point operations per second (4 Gigaflops).
• The aggregate memory capacity of the 64 v1 nodes is 256 gigabytes, allowing for extremely large problems to be run.
• The high bandwidth/low latency Giganet switch makes message-passing communication relatively efficient.

When one considers performance results, the traditional emphasis has been on processor performance, i.e., achieving a high megaflop/second rate. This is certainly a valid measure, but there are other considerations that should come into play as well.

For example, it is probably more important to have the best possible turnaround for a given job. Often this will correspond to a high megaflop/second rate, but there are examples of different code implementations where the less efficient one from a processor performance viewpoint actually runs faster (see the later discussion on algorithms).

Another important measure for academic researchers is the time to publication. While faster-running jobs aid in this process, other aspects of the computer system (I/O capabilities, visualization, mass storage, etc.) have a large impact.

At an even higher level, one can consider performance from the viewpoint of the lifetime of the application. In many cases, applications are not written from scratch but expand upon existing ones. In this case, items such as the language in which the application is written, the coding style, the power and flexibility of the data structures play a prominent role.

A certain tension exists in trying to resolve all these aspects of performance. To obtain top-flight processor performance it may be necessary to code in assembly language, making it difficult for others to use. On the other hand, the need to maintain portable software that is being used by many different people may preclude the use of some high-performance libraries. Each situation will dictate a set of compromises.