Scalable repositories for virtual clusters

For a large class of scientific data analysis applications it is becoming important, due to the sheer size of datasets, to have the option to perform the analysis directly where the data are stored, rather than on remote computational clusters. A possible strategy is the use of virtual clusters, thus guaranteeing a high degree of isolation from the underlying physical computational structure, and a very compact initial description. Deploying, saving and restoring HPC dedicated virtual clusters introduces, however, a different class of requirements on the virtual machines managing infrastructure, in particular for what concerns storage I/O requirements, whose scalability boundaries are easily reached. Here we discuss an alternative approach based on a storage model that leverages the WORM (write once, read many) character of the data used by VM management to increase, in a scalable way, the aggregate data bandwidth available to virtual cluster level operations and provide preliminary results indicating that it is a viable solution.