Section: Software

Entropy

Participants : Jean-Marc Menaud [ correspondent ] , Fabien Hermenier, Adrien Lebre, Hien Nguyen Van, Rémy Pottier, Marc Léger.

Entropy is a virtual machine manager for clusters. The current prototype acts as an infinite control loop, which performs a globally optimized placement according to cluster resource usage and scheduler objectives.

Relying on an encapsulation of jobs into VMs, Entropy enables the implementation of finer scheduling policies through cluster-wide context switches, i.e., permutations of VMs present in the cluster. It thus supports a more flexible use of cluster resources and frees end-users from the burden of dealing with time estimates.

The major advantage of the Entropy system concerns the cluster-wide context switch process itself. Entropy computes a new viable configuration and an optimized reconfiguration plan. This plan describes the sequences of transitions to perform (i.e. the run, migrate, suspend/resume, stop VM operations) in order to pass from the current situation to the new one. As the cost of each action and the dependencies between them is considered, Entropy reduces the duration of each cluster-wide context switch by performing a minimum number of actions in the most efficient way [40] .

Entropy has been presented at Supercomputing 2009, is tested by OrangeLabs, DGFiP (direction Générale des Finances Publiques), Bull, and is available under the LGPL license at http://entropy.gforge.inria.fr/ .