Dynamic Load Balancing in Real-Time Distributed Transaction Processing
Waqar Haque, Andrew Toms, Aaron Germuth · 2013
Transaction processing in distributed environments can lead to load imbalances resulting in performance degradation despite resources being available. This can have adverse consequences in a real-time environment where transactions are expected to complete by their deadlines. We propose a novel load balancing protocol, Earliest Completion Load Balancing (ECLB) that monitors the distribution of system load and network latency in an attempt to dynamically create and maintain an evenly loaded system, even in the case of heterogeneous nodes or heavy load conditions. We have assessed its performance against other load balancing protocols under normal and extreme situations. In scenarios of high load or high latency, ECLB is competitive with protocols that would otherwise be better suited to only one or the other. In heterogeneous environments, ECLB outperforms all other protocols. Among other factors, the completion of transactions before their deadlines depends upon both the network latency and load on each node. Thus, ECLB serves as a versatile load balancing protocol for realistic environments. The results are obtained using various network topologies in a parameterized distributed real-time transaction processing simulator.