A new interesting paper recently appeared online: The Impact of virtualization on Network Performance of Amazon EC2 Data Center.
The 9-pages document, produced by the Department of Computer Science at the Rice University, analyzes how hardware virtualization (EC2 is based on the Xen implementation of Red Hat Enterprise Linux) impacts on network behavior and performance, specifically focusing on processor sharing, packet delay, TCP/UDP throughput and packet loss.
The conclusions are surprising:
We have found that the networking performance between Amazon EC2 instances demonstrate very different characteristics from traditional non-virtualized clusters, such as the abnormal large delay variations and unstable TCP/UDP throughput caused by end host virtualization.
…the large delay variations can also impact many other protocols and systems that rely on the RTT measurement to infer network congestion, such as TCP vegas and PCP. Therefore, if the cloud service users want to build systems relying on the network measurements to make decisions, they need to be aware of the virtual machine scheduling characteristics of the virtualized data center environment.
Unstable network through- put and large delay variations can also have negative impact on the performance of scientific computing applications. For example, in many MPI applications, a worker has to exchange intermediate results with all the other workers before it can proceed to the next task.
If the network connections to a few workers suffer from low throughput and high delay variations, the worker has to wait for the results from the delayed workers before it can proceed. Therefore, MPI applications will experience significant performance degradation…
Thanks to @Beaker for the news.