Section: Scientific Foundations

Transparent Resource Management for Clouds

During the next years, we will continue to design resource provisioning strategies for cloud clients. Given the extremely large offer of resources by public or private clouds, users need software assistance to make provisioning decisions. Our goal is to gather our strategies into a cloud resource broker which will handle the workload of a user or of a community of users as a multi-criteria optimization problems. The notions of resource usage, scheduling, provisioning and task management have to be adapted to this new context. For example, to minimize the makespan of a DAG of tasks, usually a fixed number of resources is assumed. On IaaS clouds, the amount of resources can be provisioned at any time, and hence the scheduling problem must be redefined: the new prevalent optimization criterion is the financial cost of the computation.

Provisioning strategies

Future work includes the design of new strategies to reuse already leased resources, or switch to less powerful and cheaper resources. On one hand, some economic models proposed by cloud providers may involve a complex cost-benefit analysis for the client which we want to address. On the other hand, these economic models incur additional costs, e.g for data storage or transfer, which must be taken into account to design a comprehensive broker.

User workload analysis

Another possible extension of the capability of such a broker, is user workload analysis. Characterizing the workload may help to anticipate the resource provisioning, and hence improve the scheduling.

Experimentations

Given the very large consumption of CPU hours, the above strategies will first be tested mostly through simulation. Therefore, we will closely work with the members of the Experimental methodologies axis to co-design the cloud interface and the underlying models. Furthermore, we will assess the gap between the performances on simulation and both public and private cloud. This work will take place inside the Cloud work package of the SONGS ANR project.

HPC on clouds

Clouds are not suitable to run massive HPC applications. However, it might be interesting to use them as cheap HPC platform for occasional or one shot executions. This will be investigated with the Structuring Applications axis and in collaboration with the LabEx IRMIA and the CALVI team.