XDP2: The programming model for the datapath in AI infrastructure

We introduce XDP2– the next generation technology for the programmable datapath. XDP2 builds on the incredible success of the eXpress DataPath, or XDP. The next frontier for XDP is programmable hardware which is becoming a critical requirement for the high performance, AI/ML, datapath. XDP2 is the ubiquitous programming model for the datapath across both software and hardware. Think of it as upleveling XDP with a few tweaks and generalizations to work in domain specific programmable hardware. We’ll cover how the model defines parsers, lookup tables, accelerators, packet metadata, and inter process communications in a way that’s amenable to both software and hardware. We’ll show how XDP2 works with canonical XDP to solve the fundamental problem in kernel offloads as well as to promote XDP as a modern day microkernel.
Who is this presentation for?
Developers, programmers of AI infrastructure
Prerequisite knowledge:
Some understanding of networking
What you'll learn?
About the XDP programming model

Tom Herbert is a visionary leader in networking, software development, and operating systems with 25+ years of industry experience. He has worked at Intel, Facebook, Google, and several startups leading development of SmartNICs, networking stacks, and operating systems for data centers at hyperscale. Tom is the inventor of several industry-transformational technologies including XDP and a myriad of features in Linux networking. He has five patents granted and one application pending.