Dynamic load balancing algorithms in loosely-coupled real-time systems

Abstract

The authors study dynamic load balancing algorithms in loosely coupled hard-real-time systems. The gradient model, focused addressing and the bidding methods are used. The gradient model entails transferring backlogged tasks to nearby idle processors according to pressure gradient indirectly established by request from idle processors. The focused addressing node uses network-wide surplus information in determining the target node to send excessive tasks to. Busy nodes in the bidding method send out requests for bids to migrate tasks that are not to be completed. In the model, each job is divided into a hard task and a soft task. All hard tasks must be finished by their deadlines and will not be migrated to other nodes. If a soft task cannot be completed by its deadline, it can be migrated to a neighboring node with less load or more surplus CPU time. Three load-balancing algorithms were evaluated.<>