7.2.7 Beowulf Systems and Choices of Parallelism

		sible methods using specially crafted heuristics. Genetic algorithms typically require the execution and measurement of many programs that make up a population. This embarrassingly parallel computation is ideally suited for the loosely coupled Beowulfs.

		10.3 True Costs

		The costs of creating and maintaining a Beowulf exceed the price of its constituent components. Besides the space, power, and cooling which can be nontrivial for the larger systems, there are the personnel costs involved with the maintenance and management of Beowulfs. Beowulfs go through a three-phase life cycle: the burn-in phase, the application phase, and the upgrade phase.

		The burn-in phase involves both hardware and software. Components are most likely to fail during the first few weeks of operation. The most likely components to fail will be fans, power supplies, disk drives, and network control cards, in that order. While these parts will be under warranty, it may be desirable to initially purchase a spare so that one does not suffer extended down-times due to the delays in obtaining replacements. For a 32 node Beowulf system, a single spare would add only 3% to the total cost and provides significant additional reliability. In addition to hardware burn-in, there will be some inevitable reconfiguration of software systems as users and administrators become familiar with Beowulf operations.

		The application phase is the steady-state condition for most Beowulfs and can last for months or years. There can be very long periods of stability with system uptimes measured in months. The personnel costs incurred during this phase are related to application program development and are essentially the same as those associated with ownership of comparable commercial parallel systems.

		The third stage of the Beowulf life-cycle involves upgrades to both system software and hardware. For hardware upgrades, new nodes or components may be added, or old nodes may need to be taken off line, powered down, etc. Almost all software upgrades can be done on running systems without the need for a reboot. The only real exception is a kernel upgrade, i.e., replacing the most central component of the OS. Such upgrades are rarely needed. Linux "production" kernels are extremely stable, and rarely require upgrades. Development kernels, on the other hand, can be extremely volatile and should only be used where there is some compelling need.

		All three phases of the Beowulf life-cycle incur the costs of the personnel required to attend to the system, the physical infrastructure required to support it,