Found at ScienceDaily this week, news that UCSC scientists have created software — called PortLand — as a layer 2 data center network fabric capable of supporting more than 100,000 nodes
PortLand is fully compatible with existing hardware and routing protocols and holds promise for supporting large-scale, data center networks by increasing inherent scalability, providing baseline support for virtual machines and migration, and dramatically reducing administrative overhead. Critically, it removes the reliance on a single spanning tree, natively leveraging multipath routing and improving fault tolerance. The computer scientists report this advance in data center networking on August 18, 2009 at SIGCOMM, the premier computer networking conference.
The researchers say that PortLand gets around the routing path limitations of Layer 2 networks without the administrative overhead often associated with Layer 3 networks, especially in virtualized environments. As large datacenters, especially those in the top 25% of the Top500, grow to tens of thousands and then hundreds of thousands of nodes, developments like this could become important in simplifying HPC center architecture. The software includes some important features that support scalability, like intercepting ARPs at a local switch rather than broadcasting out to all 100,000 servers on the network.
“An important thing here is that all the switches are off the shelf—unmodified ‘merchant silicon’,” said Vahdat.
“I think PortLand is something that will be useful in the real world. The goal is to create a network fabric that allows you to buy any server or switch, plug it in and have it just work,” said Radhika Niranjan Mysore, a UC San Diego computer science graduate student and the first author on the SIGCOMM paper. Mysore presented this work at SIGCOMM 2009 in Barcelona, Spain on August 18, 2009.
A full prototype of PortLand is currently running on a network in the Department of Computer Science and Engineering at UC San Diego’s Jacobs School of Engineering.