Linux Today: Linux News On Internet Time.

IBM developerWorks: Beowulf clusters: Measuring and implementing multiple parallel CPUs

Sep 29, 2001, 13:36 (0 Talkback[s])
(Other stories by Andrew Blais)

[ Thanks to Kellie for this link. ]

"As of the middle of 2001, the world's fastest computer can perform, on average, about five trillion floating point operations per second, or 5 teraflops. The machine that ranks 500th averages about 55 gigaflops. In general, such top-tier computing power is quite expensive and unavailable (see Resources later in this article). However, in 1994, Thomas Sterling and Don Becker established that there is a way to use common and affordable hardware plus Linux to bring together the computing power of clusters of relatively smaller machines. The result, called a Beowulf cluster, can inexpensively emulate the computing power of the bottom ranks of the top machines.

The Beowulf strategy aims at minimizing computation time. A CPU operating under MS DOS can sequentially process only one program's instructions. We would like to reduce this baseline. In the case of one CPU running under Linux or NT, it is possible for multiple programs to share the CPU's resources, but all things being equal -- like word size and disk access speed -- no program processes any faster than it would under MS DOS. Moreover, in the case of Symmetric Multi-Processing (SMP) where two or four CPUs are running under Linux or NT, each processing multiple programs, it still takes approximately the same amount of time to process any single program.

Now, one option for reducing the processing time of a program is to divide it into independent sub-tasks that can be processed by different CPUs. When the results of these sub-tasks are available, they can be returned to one of the processors for final processing. Clearly, this reduces the processing time for a given program. Multi-threaded programs running under Linux on a motherboard with two or four CPUs illustrate this fact."

Complete Story

Related Stories: