Scale Up and Scale Out: Monitoring in the Data Plane

By Nic Viljoen | Mar 22, 2016

Last week at the Open Networking Summit, a key topic discussed time and time again was the need to be able to intelligently monitor network traffic in the data plane without significantly impacting performance. Tom Tofigh of the ON Lab and AT&T and Srikanth Vavilapalli of Ericsson and many others called for solutions to this problem. However, this repeatedly led to the objection that current networks would be unable to handle this level of processing in the compute nodes.

In my opinion this is just another example validating Netronome's Agilio approach to OVS offload. Having offloaded the data plane to the Agilio CX/LX intelligent server adapters for OVS or vRouter, Netronome can save server operators more than 10 cores per server at 40Gbps whilst achieving highly detailed analytics. At this point, the question of affordability is immediately raised; the Agilio CX/LX are price competitive with normal Network Interface Cards. To check out how much money we can save a particular data center network implementation read Ron Renwick’s post: "How Much Will it Cost Me?

This ability to unleash 10 or more cores in the compute node allows more complex analysis. My background is in machine learning and I believe that this is the place where very simple machine learning algorithms could add real value in networking. Being able to take some simple features from the data plane and infer higher levels of information, whether across the whole compute node or on a per VM basis, will allow not only insights at the server level, but will also allow much more targeted information to be gathered for processing at higher layers, such as in the ACORD (Analytics for Central Office ReDefined) model which was widely discussed at ONS.

Combining this with Netronome’s unique flow cache structure allows us to potentially provide data plane level, real time monitoring in a way that can scale up by being able to do more per server, and also scale out, by having an intelligent and adaptive monitoring system that scales linearly with total servers. This is a perfect example of the technical advantage that is extremely difficult to attain without Netronome’s SmartNIC approach.