Beyond Machoflops: Getting MPPs Into the Production Environment

In the continuing pursuit of the highest performance, many parallel computer manufacturers are trying to harness the promise of "killer micros" to deliver higher and higher computational rates to supercomputer users. The emphasis is on reducing clock cycles, improving hardware architectures, and increasing the number of floating point operations per second a system can deliver on benchmarks and kernels. However, there is more to a computer system than hardware. What has been lost in this pursuit is the same emphasis on providing the software that makes such systems both usable and manageable. The software fabric (libraries, utilities, programming environment, etc.) is no less important than the underlying hardware. This software aspect of a computer system is important at NCAR and many other research organizations because there are many more scientists that compute than computer scientists. A robust software environment has a tremendous impact on the ease of developing, debugging, and optimizing large codes as well as the ability to administrate such a system. Here we detail a "Top 10 List" of requirements necessary to successfully integrate massively parallel processing systems into a production computing environment, a domain currently dominated by vector supercomputers