Ciara Loftus

"With its recent advancements, AFXDP is gaining popularity in the high performance packet processing space. As a result, existing frameworks for packet processing, such as DPDK, are integrating AFXDP support to provide more options for moving packets to user space applications. The challenge with such integration is that both AF_XDP and frameworks like DPDK have their own assumptions and constraints about such things as, for example, how to align or manage packet buffers, making the integration less straight forward than it might appear at first glance.

This talk takes a look at the usability of AFXDP pre-kernel v5.4, before diving into the recent challenges we encountered when integrating DPDK and AFXDP, and how we made changes (on both sides) to allow the two to work together in a much more seamless manner."

The seamless scaling of a vSwitch in relation to changing platform workloads running in VMs/containers in real-time can be difficult, particularly when the workloads and the vSwitch are independent entities. Typically a user/orchestrator must somehow provision vSwitch cores on-the-fly or even over-provision these cores up-front, in order to ensure the vSwitch can handle the traffic going to and from these workloads at any given time. This talk introduces an approach/approaches to vSwitching that distributes the vSwitching work alongside worker cores, and as a result automatically scales with the amount of work being done on the platform. The approach is likely to be superior to existing approaches in terms of compute cost.