Abstract:
The sharing of computing resources in a distributed system is a more complex matter than in an equivalent centralised system, due to the fragmentation of resources over a set of autonomous and often physically separate hosts.
The effect of this fragmentation is reduced by evenly distributing the workload of the system over all the hosts, resulting in more effective use of the system resources, and a corresponding reduction in the average response time of processes.
This thesis concentrates on deciding how the system workload may be assigned to each host, based on the resources currently available at a host, and the resources required by the processes being distributed. Two different approaches are suggested, based on initial placement and process migration (distributing processes before and while they execute respectively). These approaches are evaluated using a trace driven distributed system simulator.
To distribute processes with initial placement requires knowledge of a process's resource requirements before it executes. Therefore the use of a simple form of prediction to provide this priori information is explored.
The major findings of this investigation are, that assigning workload to hosts based on their respective resources is worthwhile, and that process migration offers no distinct performance advantage over initial placement.
