Summary by Kyrre Begnum ...

Virtualization was a key topic at this years LISA conference with virtualization-specific tutorials nearly every day. Paul Anderson decided to run a workshop with virtual infrastructures in mind. The workshop aimed at identifying the present-day challenges in integrating and running virtualization in large infrastructures. He did a lot of work during the planning phase to get people of different fields to give short presentations. In the end Kyrre Begnum chaired the workshop.

After the presentations we did some quick polls to identify the types of attendees. The group could be divided into three general classes: One was the practitioners, who currently were using virtual machines (often in a large scale). Another group were researchers, who's main concern was management of virtual machines and automated service deployment. The third group consisted of sysadmins who were going to deploy virtualization at some point and wanted to learn more.

Most of the practitioners were using more than one virtualization technology. Everyone believed that the number of virtual machines was going to expand in the future.

The workshop was organized around short presentations with following discussions. The presentations were divided into three subjects: Deployment and Performance, Service management and Virtual machine management. The first subject was initiated by Gihan Munashinge who presented his real life experience with hosting virtual machines for customers. This talk helped set the tune for the rest of the workshop. Some very important discussion topics surfaced quickly, such as storage and lack of technology-independent management tools. Storage was by all practitioners considered a major factor in the success of the virtual infrastructure. Three dimensions of storage were discussed: reliability, performance and management. Most large infrastructures depended on redundant storage. ISCSI and NFS was common, but with a low performance in the latter. Some had created their own storage solution, such as the layered approach used in the STORM project.

Next, Lamia Youseff presented performance results from using Xen virtual machines for HPC clusters. The lack of significant performance penalties intrigued the audience, and the discussion turned towards comparing experiences and impressions on real-life performance of VMs. One interesting topic here is the way in which VMs underperform compared to a traditional hardware-based server. Performance degradation appeared to be more dramatic after a certain threshold was crossed. Lack of publications comparing performance of different technologies was discussed briefly.

Deployment issues were laid to rest and focus shifted towards deploying services  and approaches to create autonomic tools. The first presenter was Andy Gordon from Microsoft Research. The focus was on describing both the setup, but also the operational logic of a running service. A prototype system was presented, Baltic, where the overall functioning of a service was described in F#. Features such as automated scaling was supported, and could be described in operational terms. In similar lines, Nigel Edwards from HP Labs presented his experience from deploying SAP on virtual machines. He shared with us some interesting real-life issues with dynamic services and cloud-like scenarios, such as added complexity in management, software licensing and loss of control. Both presentations illustrated the potential in automated scenarios, but most practitioners used manual operations today to roll-out new VMs. For some, scripts or configuration management tools inside the virtual machine would do the individualization of the VM.

Licensing was also discussed in this context. Many licenses were VM-unaware and created problems for sysadmins. One example is licenses which are hardware-profile aware. Moving the software over to a VM from a physical server would be problematic. Also, cloning VMs would potentially violate single-copy licenses.

The last topic was management and security. Anna Fischer from HP labs talked about how to achieve secure communication between virtual machines, even when they are on different servers. Her architecture used MAC-address rewriting to create transparent communication between individual virtual machines. Several networking related problems were discussed in relation to Fischer's work, like the problem of inserting security tools into the servers in order to protect virtual machines. Further, enabling QoS on the network in order to quench VM traffic was discussed. Most practitioners did not enforce QoS on the virtual machines, instead they had several separate networks: SAN, management and LAN.

Richard Elling from Sun talked briefly about reliability and fault tolerance in virtual infrastructures. He then proceeded to talk about bundling demos into virtual machines with regard to a new storage product released by Sun.

Kyrre Begnum proceeded to talk about approaches for virtual machine management. His argument was that creating architectures for load balancing services and virtual machines were very difficult to design and saw little adoption by the community. A different approach would be to put much of the monitoring and decision making into the virtual machine itself, and letting the underlying servers play a more passive role. There was a lively discussion around this approach, where trade-offs of either approach were analyzed. Many found a so-called hybrid approach interesting, where the virtual machines assisted the decision making of the servers based on their individual policies.

This workshop provided an excellent opportunity for practitioners, researchers and curious minds to exchange ideas and experience. The discussions were fruitful with most of the 27 participants chiming in on the different subjects.

Management of virtual machines seemed to be one of the key issues for most practitioners. De-coupling the management interface from the virtualization technology would be one way in which different management approaches could be tried on the same infrastructure without switching the underlying virtualization layer. Describing the behaviour of services on a high level and transforming this description into real deployments was the challenge for the researchers. Still,  people from each camp came together in breaks and continued discussions also after the workshop. Many were interested in keeping in touch later and updating each other on new developments.