Summary by Moses Mungai & Steinar Aunsen ...

This years workshop was scheduled for a half day with 20 attendees, representing both developers, researchers and practitioners. Some present were new to the field and wanted to use the opportunity to gain insight into administration of virtual machines. We started out with a presentation from George Holmes, telling us about what he was looking for in virtualization. At his university, he planned to deploy virtual machines for students and asked for insight into process and best practices. A discussion evolved around this where it was pointed out that not all research scenarios work out on a virtualized platform, due to sometimes stringent performance requirements. However, the flexibility seems be a great advantage for the students, with the potential danger of then causing their virtual machines to be compromised or fail otherwise. This may quickly become a source of extra workload for the sysadmin, unless good backup/restore features are in place.

Next we ventured into university infrastructure virtualization, as Steve Hillman presented the effort at Simon Fraser University to move increasing numbers of servers into a virtualized platform. The motivation is based on growing confidence in virtualization as a technology as well as reduced funding and greener policies. Their total number of virtualized servers was above 200, with an ever increasing rate of migrations each month. Their framework consisted of mostly vendor products, and little self-developed tools. So far, they have had success with their approach and receive positive feedback from the departments which they host servers. One of the challenges they met was keeping inventory of the purpose of each individual virtual machines and who they belonged to.

The last presentation before the break was from Guido Trotter, who presented the ongoing work on their Ganeti framework for virtual machine management. The

system is used internally in Google and serves as a foundation for hosting virtual instances for employees and projects. Ganeti is based on Xen and KVM,

placing it in the domain of open source tools. One of the strong points of Ganeti is the ability to live migrate between servers in the same cluster.

Concurrent storage is handled by DRBD. An experienced benefit of using virtualization, they explained, is to be able to re-image virtual machines quickly

in order to recover from errors instead of hunting down the errors themselves.

After the break the workshop changed focus towards cloud computing, a topic which has enjoyed increased attention this year. Narayan Desai started out by presenting some viewpoints from the HPC community regarding cloud Computing. The most important one is the related to performance, more precisely IO latencies. Second, there was the point that commercial cloud computing may not be aimed at HPC, due to the nature of their work (sensitive data, lots of instances, high demands). Internal clouds, however, may offer some functionality towards research, since Ph.D students may deploy their experimental systems on a cloud and easily remove it when their research is finished. One point was raised during the following discussion regarding power saving: many sites have servers standing around that waste power, but no-one know what the machine is used for and therefore are afraid to shut it down and break something. Good inventory may therefore also be a source of cost reduction and powersaving.

The final presentation for the workshop was by Kyrre Begnum, who presented his research in virtual machine management. He argued, that the clouds could provide sysadmins with new opportunities if they could manage virtual machines the same way they would locally. This would become more important as the number of cloud providers increase and the system administrator would like to move its instances from one provider to another. All functionality which would be implemented in the lower virtualization layer (e.g. dynamic scaling algorithms) would be left behind when entering the new cloud. Therefore, a better approach would be to include the operational logic into the virtual machine itself and let it be its own manager. One example of such was a webserver, who could move freely between three cities based on the location of the majority of its client users.