Bring Grids Power to Internet-Users thanks to Virtualization Technologies (Virtualization)
Leaders: Adrien Lèbre (ASCOLA), Yvon Jégou (MYRIADS)
Although virtualization technologies have recently gained a lot of interest in Grid com- puting as they allow flexible resource management, Grids are still massively used through Resources Management Services (RMSs) where a static set of resources is assigned to a job (or a user) during a bounded amount of time. If one may argue that providing ded- icated time slot per job is required in some situations (e.g. in the case of reproducible experiments), a lot of end-users do not really take care about such critical requirements and just want to benefit from Grids as early as possible and as long as required (in other words, as they usually exploit their PC). A well-known approach to improve the flexi- bility of RMSs consists in exploiting job preemption, migration and dynamic allocation of resources. Each job can be proceeded, even partially, and suspended according to the scheduler objectives. The RMS performs transitions between the current situation and the expected one: jobs to stop are stopped or suspended to a disk while jobs to run are started or resumed from previously saved images. However due to the development com- plexity of preemption and migration mechanisms, advanced scheduling strategies have been rarely implemented in available RMSs. Thanks to the latest VM capabilities (such as start/stop, suspend/resume and migrate) we can consider to encapsulate each job into virtual machines (VMs) and implement finer policies through grid-wide context switches: a permutation between VMs present in the grid. Implementing such VM context switches through the whole grid requires to deal with several mechanisms from the VM image cre- ations and the related job encapsulation to the VM deployments as the supervision of the VMs until the completion of the job. If several works have been suggested to manage both VM creation and deployment, the supervision and the dynamic placement of each VM through the whole grid is still limited by several issues (such as live migration or efficient load-balancing between different sites, . . . ). By addressing such challenges, we will relieve end-users of the burden of dealing with time and resource estimates and thus we will exploit Grids in a more transparent way. Forgetting licence issues in a first step, a typical use case will consist in transparently encapsulating home jobs (for instance pic- tures or movies conversions) in VM clones of your PC and execute them into the Grid. In other words, the challenge suggested here aims at continuing to reduce the physical gap between your PC and Grids.