Intelligent Job Dispatcher for Computer Systems.

摘要

The Intelligent Job Dispatcher has two networks, the LMN-net and the information gathering network. The main result of this effort is the architecture of the LMN-net, which optimizes the assignment of N jobs to M processors in a window of L time steps. Input to this network are information on job loads, processor capacities, and inter-job dependence. Output of this network are assignment matrices. This network is a modified Hopfield network in terms of neuron connections, but its energy function contains no adjustable free parameter. The Two-Phased Optimization algorithm exploits the method of gradient descend in the 'high temperature' phase. When energy fluctuation is detected as a consequence of the combined effect of smallness of network temperature and vertex hopping network state transition, a 'quenched' phase is initiated. In this phase, the natural 'hopping-increase' in energy is exploited in the Free Mode to rescue the system from being trapped in infeasible local minima. The Free Mode alternated with the Constrained Mode which confines the system to approach a local minimum through feasible vertex points only. Simulation of application of the Two-Phased algorithm to the LMN-net for (L,M,N)=(5,3,10) and (5,6,20) are presented in this report. The purpose of the information gathering network is to assimilate the information of the input and output correlation of the LMN-net and finally replace and generalize its work in quasi-stable situations. (rrh)