Multiply-sectioned Bayesian network for multi-agent learning based meta resources scheduling in CloudSim

The interest in Cloud Computing is growing significantly. This is for the application requirements to quickly access more resources to expand as needed. Collecting real time data about Cloud resources and finding the best fit resources to user requests are time and space consuming operation. Resources can be added/removed at any time and Cloud Computing System should utilize available resources and schedule user requests within specific time for execution. Considering such dynamic and complex domain, it is important to setup global policy among resources. The global policy can be applied through resources meta-scheduling. Multiply-Sectioned Bayesian Network (MSBN) provides a natural framework for such metaphor. We propose using adaptive Multi-Agent System (MAS) based on Multiply-Sectioned Bayesian Network (MSBN) for the function of meta-scheduling in Cloud Computing Systems. Random variables are representing the uncertainties in workload and resource's attributes. Each set of identical hosts in the Cloud System are grouped as subdomains, where set of agents have partial knowledge about the global system and local observations of hosts in the subdomain. Then, agents work to estimate the state of the local hosts and act accordingly to provide probability of assigning available resources to the user request. In addition, agents need to communicate their beliefs to other agents to improve the overall meta-scheduling performance. Agent with the strongest belief will get the request assigned for local resources scheduling and execution. If the agent failed to satisfy the request needs then the agent with the next strongest belief is assigned and so on. We present our proposed model by extending CloudSim simulation. Validation of the proposed model and improvement in scheduling of Virtual Machines are presented.