A probabilistic approach to distributed batch scheduling

One of the most important objectives of the batch processing systems consists in the scheduling of processes that are to be executed overnight, or using workstations with very little or no load at all. In this way, the performance of the whole system and the individual performance of each workstation can be greatly enhanced. We propose to use Bayesian networks as a tool to specify scheduling algorithms for distributed batch processing systems. It is shown that they constitute a useful way in which the system administrator can define adequate algorithms without getting involved in the source code. The model was implemented in NQS (Network Queueing System) and it was shown that, even with very simple examples, better schedulers can be defined and their performance compare favorably with the standard schedulers provided by the system