Issue No. 03 - July-September (1999 vol. 7)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/4434.788780
This article investigates load balancing in highly dynamic local area networks connecting computers of different speeds. Most dynamic load-balancing techniques ignore the lag time in load-information update, and dispatch jobs immediately upon arrival irrespective of the overall LAN loading. This can lead to system saturation and thrashing. Moreover, these techniques focus on steady-state system throughput without considering the behavior in transient periods. So, they might not perform satisfactorily in situations that often occur in practice, where jobs arrive in clusters and system loading fluctuates widely. To tackle these problems, the authors propose two strategies: One takes network delay into account to avoid errors in scheduling jobs; the other dynamically delays job execution when the system is fully used. These strategies are general and can augment existing algorithms. Experimental results show that these strategies adapt well to load fluctuation, minimize system loading while producing shorter schedules, and improve job fairness compared to some popular schedulers.
job scheduling, load balancing, local area network, task allocation, workstation cluster
S. T. Chanson and C. Hui, "Improved Strategies for Dynamic Load Balancing," in IEEE Concurrency (out of print), vol. 7, no. , pp. 58-67, 1999.