The Linux Networking Developers (Netdev) conference is unlike any other in the industry. Conferences are usually a flurry of marketing materials, slick presentations and hyped announcements.
In contrast, Netdev mainly consists of debates about how to implement next generation features on the kernel network stack. This involves engineers from competing companies huddling in corners to ensure proposed interfaces meet their requirements, operating system distributions and software vendors dividing up implementation work between them. Interest is raised by packet metadata, not products.
And this is why I would rank this conference as the most important in the year (x2-its semi-annual) within the Linux networking space. It sets the tone for the what will be the product announcements and marketing pronouncements in a year or more’s time.
The previous conference in Tokyo was defined by possibilities: We spoke about XDP/BPF offload, many spoke about potential directions for the XDP project (named as feature of the year by Dave Miller) and there was a huge amount of news, including the oft repeated, “DPDK is *not* Linux chant.”
On the other hand, in Montreal the good news was that there was almost no significant news about new features, implementation was king. Conversations focused on using XDP for applications such as DDoS protection and load balancing, and there were talks by both FB and Cloudflare on this topic. Offloads were heavily discussed, both for BPF and also for the TC Flower subsystem, which is the focus of OVS offload currently within the kernel community.
There were significant discussions revolving around how to interact between SmartNICs and the kernel in terms of advertising capabilities to the host and the use of representors. This is great as it means SmartNICs are becoming mainstream. This fits into the themes discussed at OCP by large providers such as RackSpace.
An engineer from a competing vendor said “[we-Netronome] are clearly the leaders in this space, but others want to join and contribute.” I am excited that we are able to work with competitors on this. Not only will it allow work to be completed faster, but it will allow work to be completed in a way that suits us and others, leading to end users not having to negotiate problems such as vendor lock-in to the extent they would with non-upstream solutions.
The final comment should be left to an architect from another NIC provider: “software should not be the limiting factor, it should be about what our silicon can compete on.”
I think that is something we can all agree on.