Perf regression in SEA network provisioning
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
networking-powervm |
Fix Released
|
Undecided
|
Unassigned |
Bug Description
Deploy performance testing of current Newton code is showing a significant regression compared to Mitaka in nova's plug_vifs, which relies on the neutron sea_agent provisioning. With moderate scale (5 hosts, 400 VMs, 10 concurrent deploys), the median plug_vifs time in my tests has increased 3-4x.
My analysis suggests that the sea_agent is now generating more get_device_details / update_device_up cycles due to the recent removal of the PVIDLooper, because the CNAEventHandler is now triggering more provisioning operations which used to be coalesced by PVIDLooper. These additional port status changes drive notifications across neutron and nova on both the host and controller, and this extra activity is getting in the way of the actual deploy operations.
description: | updated |
Fix proposed to branch: stable/newton /review. openstack. org/385592
Review: https:/