A Semi-Distributed Load Balancing Model for Parallel Real-time Systems.
K. Erciyes, Öznur Özkasap, Nilgun Aktas · 1995
We propose static and dynamic load balancing policies for parallel real-time systems. A parallel real-time system in this context is considered as a computational environment consisting of a number of processors where stringent timing requirements of processes should be met. This would encompass massively parallel systems at one end of the spectrum and a group of computers connected by a local network at the other end. The static and dynamic load balancing policies developed are suitable for both types of systems with parameters such as communication costs to be tuned for each environment. For massively parallel processing systems, we introduce the concept of a domain which is a pool of processors and is governed locally for various services such as dynamic load balancing. The dynamic load balancing is implemented by central load balancers per domain which make use of the group communication facility for distributed communication with the other load balancers. This semi-distributed approach eliminates the need for maintaining a central node or replicated data by providing local data and control confined to that domain. The distributed data and control transfer is performed among the servers of the domains. The static scheduler however works offline for tasks with known characteristics such as execution time, communication constraints and deadlines prior to their execution which would be the usual case for hard real-time tasks.