Scheduling highly available applications on cloud environments

Cloud computing is becoming a popular solution for storing data and executing applications due to its on-demand pay-per-use policy that allows access to virtually unlimited resources. In this frame applications such as those oriented towards Web 2.0 begin to be migrated on cloud systems. Web 2.0 applications are usually composed of several components that run indefinitely and need to be available to end users throughout their execution life cycle. Their availability strongly depends on the number of resource failures and on the variation in user hit rate. These problems are usually solved through scaling. A scaled application can span its components on several nodes. Hence if one or more nodes fail it could become unavailable. Therefore we require a method of ensuring the application's functionality despite the number of node failures. In this paper we propose to build highly available applications, i.e., systems with low downtimes, by taking advantage of the component based architecture and of the application scaling property. We present a solution to finding the optimal number of component types needed on nodes so that every type is present on every allocated node. Furthermore nodes cannot exceed a maximum threshold and the total running cost of the applications needs to be minimized. A sub-optimal solution is also given. Both solutions rely on genetic algorithms to achieve their goals. The efficiency of the suboptimal algorithm is studied with respect to its success rate, i.e., probability of the schedule to provide highly available applications in case all but one node fail. Tests performed on the sub-optimal algorithm in terms of node load, closeness to the optimal solution and success rate prove the algorithm's efficiency. (C) 2012 Elsevier B.V. All rights reserved.