Article Preview
TopIntroduction
Resource capacity planning is becoming a big challenge for cloud providers when modern applications require a huge amount of resources. To adapt with the environments where resources are dynamically requested and released, conventional approaches that overprovision VMs with a large amount of resources to meet these requirements can induce several problems such as underutilization of these resources in most cases. Oversubscription is an adequate technique that deploys more VMs as much as possible. This technique becomes feasible since customers tend to overestimate their resource requirements. It can alleviate the problem of resource underutilization which allows minimizing resource provisioning costs while meeting customers’ needs. This technique has inevitably some risks: it can increase the possibility of overload especially when the oversubscriptions ratios are not managed carefully.
In this paper, we present an autonomous architecture that improves the resource utilization rate through the technique of memory oversubscription. The overload induced by this technique is mitigated by two techniques: Live Migration and Network Memory. The first strategy consists of live migrating some VMs from an overloaded host to an underloaded one without interrupting any one. The latter consists on using the global memory of the cluster as a remote paging replacing the disk swapping as it has a best performance (nano second vs. micro second) (Baset, Wang, & Tang, 2012).
Live migration is a technique largely used to improve service performances in an IaaS Cloud. It provides significant benefits, such as minimizing the Service Level Agreement (SLA) violations and insuring the quality of service (QoS). Although several live migration algorithms had been implemented, they do not address the migration cost. Here, our proposal aims to optimize the migration cost which is an important concern.
On the other side, the protocol RDMA (Remote Direct Memory Access) is used to mitigate the overload due to oversubscription if the network memory strategy is decided. This protocol allows to transfer data without involving the CPU or cache for both source and target hosts.
For the oversubscription ratio, unlike traditional strategies using, either static ratios or depending only of the local host capacity, we adopt a dynamic oversubscription one, taking into account the free memory of the local host as well as that available on remote hosts of the cluster. This ratio is dynamically managed by reporting the amount of oversubscribed memory in real time.
Simulation results show the effectiveness of our proposal -in term of improving service performance- compared with the use of live migration alone as a technique to mitigate overload.
Our contribution consists essentially on:
- •
Maximizing resource utilization rate in the cluster with the technique of oversubscription by adopting a real time and dynamic oversubscription ratio;
- •
Improving service performance by leveraging the global memory of the cluster trough the network as a technique to mitigate the overload at any host via the remote paging technique.
The remainder of this paper is organized as follows: In Section 2, a brief description of oversubscription is introduced in the context of Cloud Computing. Section 3 describes some related work in the field of oversubscription and overload mitigation strategies. In section 4 we propose an autonomous architecture that mitigates memory overload by the two techniques described above. To validate our proposed approach, we present our simulation study in section 5. Finally, in Section 6 we conclude our paper by specifying some future work.