Transparent load sharing in distributed systems: decentralized design alternatives based on the Condor package

In recent years a number of load sharing (LS) mechanisms have been proposed or implemented to fully utilize system resources. We design and implement a decentralized LS mechanism based on the Condor package, and give in this paper a description of our design and implementation approaches. Two important features of the design are the use of region-change broadcasts in the information policy to provide each workstation with timely state information at the minimum communication cost, and the use of preferred list in the location policy to avoid task collisions. With these two features, we remove the central manager workstation in Condor, configure its functionalities into each participating workstation, and thus enhance the capability to tolerate single workstation failure and the reliability of Condor. We also discuss the experiments we conduct on the LS mechanism and the observations we obtained from empirical data.<
>