Optimizing Logical Structures of Decentralized Schedulers in Multicluster Systems and Computational Grids

In this work, we study the influence of local neighborhood choice to the efficiency of decentralized scheduling in geographically distributed multicluster systems and computational grids. We constructed the mathematical model for evaluation of the efficiency of the logical structures. We also designed efficient algorithm, which builds suboptimal graphs of computer systems' local structures. We give the results of full-scale experiments of scheduling of MPI-programs using common logical structures, which demonstrate the influence of the local neighborhood choice to the efficiency metrics. We also compare common structures with the structures, generated by means of designed algorithm. The results of numerical modeling show that the proposed algorithm significantly reduces the cost of job service.