Video recording and production done by OpenStack Foundation.
Neutron has a good feedback cycle between getting new features added and getting them tested by other contributors so they are functionally correct within one or two releases. Unfortunately, after a feature is working ("it worked in devstack!"), developers tend to move on to the next new thing. This happens well before the previous features are deployed at a large scale.
This delay has led to limited attention being paid to the performance of Neutron's HTTP management API and the AMQP control plane it uses to communicate with its agents. Subsequently, several cases had emerged over the past few cycles where operations were taking orders of magnitude longer than they should have to complete. These were imperceptible to developers because they were usually amplified by large numbers of networks and ports (instances) not present in a typical development environment.
This presentation will cover the wide range of performance improvements made to the Neutron management and control plane over the last ~6 months. These include both user-facing improvements (e.g. the time it takes to list Neutron networks) and deployer-facing improvements (e.g. the ratio of Neutron servers to L3 agents required to respond all agent requests). Benchmark and improvement numbers will be provided using measurements taken by Rally. This will be focused on the open source ML2 reference implementation; however, most of these improvements benefited many 3rd party plugins/drivers that utilize the same APIs.