Adaptive memory load management in cloud data centers

摘要

We consider a cloud data center in which a set of application servers is hosted. Each server runs in a virtual machine in the cloud, subjected to a session-oriented workload, whereby session data of one server is replicated on other servers for purposes of high availability. In such an environment, we are concerned with the memory resource usage due to the application servers. In particular, our goal is to prevent memory overload by managing the session load admitted to each of the application servers. Little interest has been given to load management in the cloud based on memory usage, although memory is a crucial and valuable resource. We introduce a dynamic memory overload protection solution that is based on adaptive feedback controller techniques. In particular, we have designed and implemented a self-configurable memory controller, which is automatically tuned based on an analytical model of the system under control. Our proposed solution consists of a set of independent controllers and, hence, is a scalable architecture. Challenged by actual correlation among application servers due to session replication, we validate our solution on commercially available servers in the cloud environment. Our experimental results illustrate good performance in the presence of load fluctuations and session replication.