Efficient Load Balancing on a Cluster for Large Scale Online Video Surveillance

In this paper we present a new load distribution strategy tailored to real-time, large scale surveillance systems with the objective of providing best effort timeliness of on-line automated video analysis on a cluster of compute nodes. We propose a novel approach to fine grained load balancing, modeled as a makespan minimization problem to reactively minimize the tardiness of processing individual camera feeds. The proposed approach is also robust in the sense that it is not dependent on either the estimates of future loads or the worst case execution requirements of the video processing load. Simulation results with real-life video surveillance data establish that for a desired timeliness in processing the data, our approach reduces the number of compute nodes by a factor of two, compared to systems without the load migration heuristics.