Minimizing Geo-Distributed Interactive Service Cost With Multiple Cloud Service Providers

Geo-distributed interactive service is a type of Internet cloud service. Since data centers belonging to multiple cloud service providers (CSPs) can provide lower resource provision cost for these services, we try to overcome the two challenges in selecting the data centers from multi-CSPs for our cost optimization. First, as the resource prices vary across data centers, scheduling data centers and geo-distributed interactive services have a tremendous consequence for optimizing the cost. Second, the resource demand is uncertain, and the geo-distributed interactive service needs enough resources to meet the quality-of-service constraint. A regularized stochastic decomposition algorithm, namely, two-dimension resource provision (2DRP), is proposed for solving the two-stage stochastic linear programming problems. In stage one, users express tail latency constraints for each geo-distributed interactive service, letting 2DRP choose the lowest price-per-instance data center group and determine the right amount of resources. In stage two, 2DRP estimates the impact of the request rate and the amount of resources on tail latency and uses the decomposition technique to adjust the resource provision plan to handle the demand uncertainty. We evaluate 2DRP over a set of workload scenarios on real clouds by reproducing the Internet request with data center workload traces, and the results show the availability of our algorithm for cost minimization geo-distributed interactive services with multi-CSPs.