Case Study: Clusters and Image Processing, Part I
In 1998, Mark Lucas and the other folks at ImageLinks, Inc., already had a product. Forty programmers and scientists had spent 14 years building the original version for the Harris Corporation. This software was classified during the cold war and finally declassified in the early 1990s. ImageLinks spun off of Harris to apply this same technology in the commercial remote sensing market, which involves methods used to get information about the earth from somewhere other than the earth's surface.
This program is a collection of tools that work together to provide image analysis. Pictures taken by remote sensing equipment--typically from a satellite or a digital camera on a high-flying airplane--are fed into the program, mapped onto a three-dimensional model of the earth, and manipulated, merged with other images from other equipment, or who knows what else.
In the early days of ImageLinks, the software ran on a network of a dozen SGI Indigo 2 systems with R10000 processors. This network was managed over Asynchronous Transfer Mode (ATM), a networking technology that can achieve data-transfer rates of 10Gbps, using an SGI R8000 four-processor Challenge server. This collection of processors and networking was extremely high-end equipment for the time. ImageLinks had to have a computing powerhouse to accomplish such processor- and bandwidth-intensive tasks.
NOTE: An R10000 CPU is one in a series of 64-bit, MIPS-class processors. This level of CPU is traditionally used in high-end servers and mainframes. MIPS processors are also referred to as superscalar because they can handle multiple instructions simultaneously from start to finish.
Mark and some of the other technical folk at ImageLinks were experimenting with Linux at home. They began wondering how the application would run on a set of Linux boxes instead of on the SGIs. Their theory was that they would not make much of a gain. However, if the result was comparable, a move to Linux could save them thousands of dollars a year on software and licenses.
There was only one way to find out. They had to give it a try.