Group-based memory oversubscription for virtualized clouds

As memory resource is a primary inhibitor of oversubscribing data centers in virtualized clouds, efficient memory management has been more appealing to public cloud providers. Although memory oversubscription improves overall memory efficiency, existing schemes lack isolation support, which is crucial for clouds to provide pay-per-use services on multi-tenant resource pools. This paper presents groupbased memory oversubscription that confines both mechanism and policy of memory oversubscription into a group of virtualmachines. A group is specified as one of service level agreements so that a cloud customer can control the memory management mechanism within its own isolated domain. We introduce group-based memory deduplication and reprovisioning with several policies based on per-group workload behaviors. The proposed scheme is implemented on the KVM-based prototype and evaluated with realistic cloud workloads such as MapReduce and MPI applications. The evaluation results show that our group-based memory oversubscription ensures strict inter-group isolation while achieving intra-group memory efficiency, compared to a system-wide scheme, by adapting oversubscription policies based on per-group workload characteristics. © 2014 Elsevier Inc. All rights reserved.