Arista presented at Networking Field Day 16 about higher speeds, embedded customer code, streaming analytics, and APIs. Arista introduced itself as focusing on IP storage and big data, using deep buffers end-to-end. But, as we’ve seen with some TCP stacks, deep buffers can cause problems. (Read about a buffer problem in the blog post: What a network speed check revealed about sluggish performance.)
Arista Founder, Chief Development Officer, and Chairman Andy Bechtolsheim talked about 400G Ethernet, explaining how the industry will get there. He predicted 50G links in 2018 as an interim step. A 100G-per-lambda specification is in development, to be available in 2020. Andy has always been an excellent hardware designer, and he shared with us that 100G for the electrical lanes is about as far as technology can go. I wonder about an implication that we’ll have to look to pure optical to go faster.
Founder and Chief Technology Officer Ken Duda spoke about the ability for customers to run their own software on Arista switches, using a stock Linux OS. This creates some interesting possibilities regarding high-speed access to the state of the switch. Customer code is decoupled from Arista code and a published database provides the interface between the two.
Ryan Madsen, software engineer, then told us about Arista’s streaming analytics and monitoring system. A customer question about this type of technology resulted in a recent blog post about the data-driven network. In Arista’s case, the data are sent to their analysis engine, CloudVision, which looks like a publish-subscribe (pub-sub) message queue system. A customer application and an API are available to allow customer access to the analysis results. The current GUI wasn’t particularly impressive, but this is a first version, and I expect them to get customer feedback and make improvements.
Like most vendors in the industry, Arista provides an API for control and monitoring. Its JSON interface corresponds to its CLI command set, with the caveat that some models may not be able to reflect exactly everything that the hardware is capable of doing. Like the streaming analytics, they are working toward using CloudVision to perform system-wide configuration versus box-by-box configuration.
I think it is interesting to see vendors starting to build mechanisms to export data from network devices. SNMP may be looking at its replacement. But we need a standard data export format before SNMP will be replaced. If you’re thinking JSON is the answer, remember that it is only an encoding mechanism that will need to be extended if it is going to replace SNMP. For more thoughts on these mechanisms, read my blog post, “Is the Data-Driven Network the Next Step in Networking?”
Disclosure and caveat: While the NFD presenters indirectly paid for my travel and lodging to attend NFD16, I didn’t receive any compensation nor have the vendors been promised any favored treatment in articles or reviews. I’ve not used CloudVision myself, so my opinion on it is simply from the brief presentation at NFD16.