On-the-fly Networking for Big Data
"Big Data" is one of today's most interesting usage scenarios for data centres. Huge amounts of data have to be stored, processed, and coalesced on many distributed computers. This poses a significant challenge for the infrastructure of even big data centers, specifically, for the data center networks. Specifically, it becomes necessary to custom-tailor the network's topology to the particular application mix running in the data center and the traffic patterns it entails. Such a custom-tailoring has to happen at runtime since the application mix can and often will change in an unpredictable way; therefore, it has to happen in software since rewiring is not an option. Moreover, the network behavior overall, not just the topology, need to be flexible.
One answer to this need for flexibility is the introduction of so-called Software-Defined Networks: Networks where the switches only provide the smallest possible functionality (i.e., packet forwarding) but delegate all decision logic to remote controllers. While that is conceptually a nice idea, it raises considerable scalability and performance issues.
This project group will set out to address some of these concerns. We will develop
- a concept how controller logic can be split up in simpler pieces to be executed on the network switches
- complex logic that still remains at the remote controller
- ways to put these simpler control codes into an execution fabric on a network switch
- realize a testbed to evaluate our solutions.
To this end, we will leverage an existing SDN testbed based on NetFPGA cards with 10 gbit/s Ethernet equipment. These FPGAs can be used to execute simple control logic by being reconfigured on the fly once such code is installed in a network.
This project group is jointly organized by the Computer Engineering group (Prof. Platzner) and the Computer Networks group (Prof. Karl) and takes place in the context of the SFB 901.