Spider Mesh Overlay for Task Load Balancing in Cloud Computing

With the growth of the Cloud Computing paradigm, a new challenge appears in the development of data centers: energy efficient load balancing in a dynamic and scalable environment with an undefined number of heterogeneous resources. In this context, considering that idle servers consume up to 60% of peak energy and the fact that usually the level of usage for a server is 20-30%, new load balancing strategies are required. In this paper we present the various properties of a novel, biologically inspired, spider mesh overlay used for load balancing non-preemptive tasks, each task consuming a certain amount of resources. Each node of the overlay is capable of offering the best capacity for a certain resource (e.g. powerful CPU for CPU-bound tasks) while offering smaller capacities for the other resources. Various routing policies are used in order to assign a job to a fitting node or in the worst case, provision/switch-on another node. The jobs are distributed by a central broker using two methods: round-robin and a clustering approach. In order to test our methods we implemented a simulator for the datacenter and the spider mesh overlay, the simulated resources being sampled from a gaussian distribution. Experimental results are promissing as both the round-robin and clustering approach show an overall server load of more than 50% in the proposed worst case scenario.