DPDK PMD Live Upgrade

on YouTube ↗25m

" summary

Adds an API that lets a new mlx5 PMD start in active/standby mode so flow rules and configs can be copied while the old PMD still forwards traffic, then flip the new PMD to active by deleting a temporary group zero dispatcher entry. Claims full flow-offload migration with zero downtime for the flow plane, supports optional simultaneous mirroring to old and new, scales to user hardware limits, is validated with two CSPs, and currently targets Mellanox mlx5 (proposal to generalize to other PMDs remains open).

AI-written summary · about

" chapters (6)

1. 0:00Introduction to PMD live upgrade
2. 6:48Upgrade requirements and goals
3. 10:27New upgrade sequence with active standby
4. 14:03Extending active standby beyond flow engine
5. 18:21Use cases and future extensions
6. 21:12Q&A on upgrade process

" related talks

• 15mKVM Live Upgrade with Properly Handling of Passthrough DevicesZhiming Feng · KVM Forum 2020
• 14mDeep Optimization of VMM Live UpgradeShenming Lu · KVM Forum 2022
• 50mOps issuesChris Morgan, Erik McCormick · OpenInfra Summit Virtual 2020 Forum
• 27mFailover Without Fear: A Practical Guide to Active-Active...Sina Moghaddas · SREcon25 Europe/Middle East/Africa
• 31mProgramming the Performance Co-Pilot toolkitNathan Scott · PyCon Australia 2014
• 46mEnd-Users Won't Even Notice, Painless Kubernetes UpgradesRicardo Aravena · All Things Open 2020
• 24mSupport for Fast and Reliable VMM Live Upgrades in LibvirtSoham Ghosh, Prachatos Mitra · KVM Forum 2021
• 17mPIMS: Fleet-Wide Datacenter Maintenance with Minimal Capacity Buffer and Predictable Lat…Ben Christensen, Kaushik Veeraraghavan, Benjamin Leonhardi, Yang Wang, Mayank Pundir, Evangelia Kalyvianaki, Abdelrahman Adam, Agshin Nabiyev, Aleks Shirokov, Amitav Mohanty, Daniil Balenko, Elaine Zhao, Essam Ewaisha, Hongbo Dong, Igor Marnat, Lev Novikov, Min Zeng, Steven Shingler, Timofey Durakov, Wiliam de Abreu Pinho · 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26)