Market-based autonomous and elastic application execution on clouds. (Gestion autonome des ressources et des applications dans un nuage informatique selon une approche fondée sur un marché)

ions in the management of physical infrastructures gives more flexibility to users in controlling the environment in which their applications run. Given the recent efforts to improve current virtualization mechanisms and better support HPC workloads, i.e., by virtualizing specialized network devices like Infiniband [72], using virtualization in HPC centers becomes feasible. In this section we summarize the resource management solutions that used virtual clusters. Haizea Haizea [157] is a centralized lease-based resource management system for VMs. A lease is defined as a set of VMs made available to a user based on a contract that specifies the provisioning terms, like VM resource requirements and the lease duration. Haizea supports different types of leases, from best-effort to advance reservation, and handles the overheads related to VM management. By scheduling the VM deployment before the start time of the lease, the system ensures that the time required for VM management is not charged from the lease period and, thus, it does not delay the execution of jobs for which resources were reserved in advance. This model uses the scheduling algorithms implemented by batch schedulers. Maestro-VC Maestro-VC [96] is a partially decentralized solution that creates and manages dynamic virtual clusters. This solution relies on a global scheduler that decides the resource allocation for each virtual cluster and local schedulers, running in the virtual clusters, which negotiate their virtual cluster allocation with the global scheduler. Thus, the local scheduler can take resizing decisions, allowing the application to run with less resources, or it can checkpoint the application, thus avoiding loosing the already performed computation. If resources need to be revoked from a virtual cluster, the global scheduler notifies the corresponding local scheduler. Unfortunately, this solution does not include policies to decide the resource amount each virtual cluster receives. Violins Violins [146] are autonomic virtual clusters that are resized and live migrated between nodes, and possibly different physical domains, to meet a certain application resource utilization level. Each application from the system runs in its own VIOLIN. S tarting from the assumption that the application resource demand can be dynamic inside the VM, a VIOLIN can be resized in a fine-grained manner. The resource allocation decisions are taken by a global adaptation manager, which retrieves resource utilization from physical nodes and VMs, and decides the resource allocation and placement of each VIOLIN to satisfy the application resource demand. This solution is actually a centralized system, as the main allocation decisions are taken by a central entity, based only on VM utilization metrics. Moreover, they do not consider per-application performance objectives. ORCA ORCA [76] is a middleware solution for federated domains that relies on three entities: consumers, providers and brokers. As illustrated in Figure 2.3, these entities interact with each other to schedule and share resources. Consumers, which are actually application agents, request and acquire tickets from brokers, entities that aggregate and manage resources from one or multiple providers. In order to actually receive the resources, the consumer needs to present the ticket to the provider, which in turn leases the resources to it. Basically, tickets represent the ownership of the acquired resources, and specify, among others, the identity of the provider to which those resources belong. The lease, instead, represents the actual resource allocation. This resource model is attractive for federated environments as it allows providers to keep a level of control over their resources. The way leasing is implemented in ORCA allows