Scaling considerations

Abstract

This chapter presented scaling considerations taken in MOSIX in order to allow the system to function properly when used in a large multicomputer configuration. It identified some of the difficulties associated with using existing operating systems mechanisms, and then defined a set of principles for constructing algorithms for large scale multicomputers. These principles include symmetry, client/server protocols, partiality, and randomness.

To demonstrate the effectiveness of these principles, the chapter presented several probabilistic algorithms for collecting information on the state of the system, for use in management and control of resources. The interesting properties of these algorithms are their completely decentralized nature, their use of only partial information about the global state of the system, and their limited degree of node coordination. The performance of the algorithms in terms of the quality of results and overhead per processor is insensitive to the size of the system.

Alternative algorithms for disseminating information reliably without broadcasting in a LAN-based system are described in [1]. These algorithms use repeated forwarding and packaging of messages to disseminate information. In these algorithms, broadcast of a message to all the processors can be achieved quickly with only point-to-point messages.