Advances in switching are achieved with the OpenFlow switch architecture from Stanford University. The concept separation of switch fabric and switch control is also playing a major role. It allows for cheap switches, centrally controlled by the switch manager. The assessment is suitable for low-latency data centre communication.

Advances by virtualisation and parallelization are characterized by disruptive technologies such as virtualisation and multicore architectures in the network. The drivers for virtualisation are the complexity which virtualisation allows to hide, and new management principles.

Prof. Dr. Carle also talked about the advances in communication software which can improve wire-speed packet capture and transmission. Linux has developed the PF_ Ring for network socket for high-speed packet capturing. New as well is NAPI, new API interrupt mitigation techniques for networking devices in the Linux kernel.

Software issues have to be dealt with as well, such as crucial software for low-latency.

Prof. Dr. Carle also had a lot of questions. As far as the Cloud is concerned, how suitable are Cloud concepts for HPC applications. Are there any limitations? And what is the typical availability? Which overhead does virtualisation impose? Can we afford it in HPC applications?

Prof. Dr. Dieter Kranzlemüller from the Leibniz Supercomputng Center, one of the three national supercomputing centres in Germany, tried to formulate an answer some of these issues as the first speaker in the session. He told the audience that the Leibniz Center is providing Cloud but they do not call it Cloud as such.

In fact: what is Cloud? And what are the differences between Grids and Clouds? Which technologies can be run in a Cloud? Prof. Dr. Kranzlemüller cited a commercial definition of Cloud, saying that Cloud is on-demand access to virtualized IT resources but this is a non-scientific definition.

He recalled to the audience that Thomas Lippert was defending HPC against Cloud at the ISC'09 Conference last year. Thomas Lippert then said that HPC actively offers the highest level support and research for science communities and industry and that it can offer leading-edge tier-3,2,1,0 HP systems.

So Thomas Lippert thinks that Cloud is insufficient.

Prof. Dr. Kranzlemüller also referred to a benchmarking effort for Amazon EC2 by Edward Walker in October 2008.

So what are the commuication aspects for MPI bandwidth? And what about the comparison between Cloud and cluster? There are large difference in bandwidth between the two, showed Prof. Dr. Kranzlemüller. HPC is not so optimal as people want to believe.

On the other hand, if Cloud computing centres do not use higher bandwidths there are no results of improvement. Two main issues are access to CPU and memory access. These are important factors to be taken into account because the performance depends on it.

There are different parameters, such as the application, virtual machines, the hypervisor, and the hardware. To start the benchmarking, one has to distinguish between the static and the dynamic parameters. The dynamic parameters are observed during the run time.

Theoretical performance calculation is not possible, so benchmarking is needed. The benchmarking is being done on different platforms.

In sum, there is a constant throughput and an equal distibution among the virtual machines.

There will be a collapse of the reading throughput, if the number of virtual machines is larger than the number of CPUs. This will cause scheduling issues.

The virtualization of existing infrastructures calls for capacity-planning. The virtualization of the CPU and the main memory is highly efficient, nearly 100%, according to Prof. Dr. Kranzlmüller. The operation of virtual infrastructures depends on the distribution of virtual machines among the servers.

The type of virtualization and implementation is crucial for high performance computing. The distribution of virtual machines among the hosts is very important. If there is a lack of knowledge about the underlying virtual topology, it will be hard to fine-tune the parallel applications.

Further challenges involve the security of the virtual machines and the hosts, the data and the users, Prof. Dr. Kranzlemüller concluded.