DISTRIBUTED AND DECENTRALIZED CONTROL IN FULLY DISTRIBUTED PROCESSING SYSTEM

An essential component of a Fully Distributed Processing System (FDPS) is the distributed and decentralized control. This component unifies the management of the resources of the FDPS and provides system transparency to the user. In this dissertation the problems of distributed and decentralized control are analyzed resulting in the construction of four models of control which are analyzed by means of simulation experiments.;A partitioned management strategy enlisting a global search in order to locate resources for a user request is utilized in the first control mode. The second model of control maintains a central directory of all resources. The third model is similar to the first model with the exception that resources are located by first searching locally and only conducting a global search if all resources could not be found locally. The fourth model of control maintains identical, redundant resource directories on all nodes with access to the directories provided in a serial fashion by passing a special message called the control vector among the nodes. Modifications made to a directory by the holder of the control vector are transmitted to all other nodes.;Four groups of simulation experiments were conducted. The first group of experiments examined the behavior of jobs accessing local files while the second group investigated the behavior of jobs remotely accessing files. Both experiments indicated little difference in the performance of the various models. In the first group the average work request response times approached a constant value similar to that obtained with a single node network. The third group of experiments, which studied jobs not requiring file access and possessing small service times, provided a clear distinction among the models with a relative ordering from smallest to largest average work request response time as follows model 2, model 3 model 1, model 4. A mixed population consisting of the jobs from the second and third group of experiments was analyzed in the fourth group of experiments. These experiments indicated that the average work request response times for the short jobs increase as the fraction of jobs accessing remote files was increased.