Maximizing Cloud Providers' Revenue

With the increasing number of jobs executed and accomplished in clouds,more and more data centers are deployed and cloud providers generate revenue by leasing out data center resources to multiple tenants.A serious problem faced by those providers is how to maximize their revenue under guaranteeing the quality of jobs submitted by tenants.Previous works mainly focus on improving the resource utilization by assigning tasks to each hosts reasonably,which increases providers' revenue in an indirect manner.However,most of them are offline approaches which cannot be practical in clouds,because not all the information of tasks can be known,especially the ones that have not arrived and their information can only be known at their arrival.This thesis firstly abstract the jobs of tenants and data centers into graphs,where information of nodes contains CPU and storage resources and the weight of edges quantify the bandwidth requirement between two virtual machines(VM)or capacity of links.Then,we model the process of achieving the approximately optimum solution as an invertible Markov chain,in which each flexible solution is regarded as a configure and each configure can be reached from others according to relevant probability.By KKT condition,we derive an important conclusion of the transport probability between two configures.In addition,the paper considers constrains of multiple resources and explores the(1-1/e)-competitive online algorithm using primal-dual technique.At last,the accuracy of theoretical lowbound of that online approaches is demonstrated by the simulation experiments.