The Programmable Data Plane and the OS

2022-02-22 Draft ~ / architecture /

Background

Over the last decade or so, data networks have become increasingly programmable. Fixed function network interface cards and switch ASICs – once the only option – are now readily available in various programmable forms. Much of the work in this space, sometimes referred to “software defined networking” has focused on the development of new protocols, often skipping the operating system's networking stack all together.

The programmable data plane also offers potential advantages to the OS network stack. First and foremost, network hardware is no longer a black box. This means that interactions between kernel-level network functions and network functions exposed by hardware can be analyzed and understood as a combined whole. A programmable data plane also democratizes development of kernel-level network features by putting a community of kernel developers in the driver seat rather than living at the mercy of proprietary firmware.

A formidable challenge in incorporating a programmable data plane at the OS level is supporting the large breadth of network protocols that are required. From servers, to laptops to embedded devices – the number and variety of protocols that must work together for the operating system's networking stack to function effectively presents a different set of challenges than incorporating a programmable data plane for one-off kernel-bypass programs.

Network Functions at the Hardware/Software Interface

The following network functions are a few simple examples of how OS-level and firmware-level network functions interact. In this post we'll refer to firmware simply as code that executes on a network device independently of the operating system. The examples presented are nowhere near an exhaustive list. Take a look at the data sheet for any modern data-center NIC 1 2 3, and you'll find dozens of network functions that are implemented by these devices.

Checksum Offload

TCP Offload

Tunnel Encap/Decap Offload

Pushing more functionality onto network processing units

Ideally, the central processing unit on a computer is dedicated to executing the code of programs running on the computer, and spends few cycles as possible performing network functions. This leaves the execution of network functions to network processing units (NPU). However, for some network functions, the OS must be aware of a subset of the state involved and how it evolves. This is called stateful offload. Functions that can be completely offloaded to an NPU and the OS can remain happily oblivious to what's going are referred to as stateless offload.

The Fragmented state of Data Plane Compilers and Control Interfaces

An OS-Level Programmable Data Plane Architecture

Working Backwards from the Ideal Interface