Performance optimization of Neutron server and Agent for large-scale scenario
summary
500-node Rocky-based scale test reveals DHCP and OVS-agent bottlenecks; DHCP agent accumulates message backlogs causing nova port activation timeouts, and OVS agent spawns many conntrack delete processes that steal CPU. Solutions: prioritize and batch same-network port messages and add a message-watcher to pull and handle network-wide messages in bulk, and replace concurrent per-port conntrack deletes with a single contractor-driven delete worker. Results: stem-network message delays drop from ~6 minutes to seconds/minutes at 10–30 msgs/s, and conntrack deletion for 60 VMs drops from >60 minutes to <5 minutes (12x improvement) while keeping nova’s 300s wait. DHCP batching single-process conntrack deletion
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:13Large-scale test overview
- 4:11Dhcp agent message backlog
- 9:51System network message bottleneck
- 13:12Optimization testing results
- 17:47Layer 2 agent performance issues
- 22:26Os agent connection tracking fix
- 27:02Performance improvement results
related talks
36mMigrating the Neutron backend to OVN in production
35mWhy OCP networking performances on VMs are suboptimal and how to improve them by 5x
01h 16mTroubleshooting OpenStack Neutron with ML2/OVN backend driver
30mNoisy Neighbors, through Networking
32mExtending Network Quality of Service support in OpenStack
23mOptimizing Live-Migration for Minimizing Packet Loss With KubeVirt and...
18mOpenStack in Telco: How to achieve Low latency computing with OpenStack
20mCan OVS-DPDK be further optimised?