router_info process_floating_ip_addresses execution time is O(n)

[Ryan Moats]

router_info's process_floating_ip_addresses execution time increases as the number of routers scheduled to a network node increases. Ideally, this execution time should be O(1) if possible.

[Ryan Moats (rmoats)]

Using 3.19 kernel appears to address this, so marking invalid

[Assaf Muller (amuller)]

Can you explain how the kernel version has anything to do with the bug reported?

[Tapio Tallgren (tapio-tallgren)]

The function process_floating_ip_addresses is a methon in router_info, so I do not see how the processing time would increase as a function of #routers. It seems to be O(n) on the number of cidrs configured on a router (since it has a set of of cidrs and checks membership in the set). Is this the meaning?

[Ryan Moats (rmoats)]

@tapio-tallgren, @amuller: it turns out that this code invokes ip netns exec to perform some operations. In earlier kernels (for example 3.13) that command's execution time increase with the number of namespaces that have been created. In more recent kernels (3.19 for example) this behavior has been fixed so that the execution time does not increase with the number of namespaces that have been created.

@tapio-tallgren: Since namespaces is equivalent to routers, that's where the O(n) on routers comes from

@amuller: Once I ran with 3.19 kernel, the execution time became O(1) so I marked the bug invalid.

[Assaf Muller (amuller)]

You could convert this knowledge of kernel versions in to a doc enhancement. Which doc, I wouldn't know, but we can ask around :)

[Ryan Moats (rmoats)]

@amuller: After talking to mestery last week, I added the following section to the Liberty release notes:

https://wiki.openstack.org/wiki/ReleaseNotes/Liberty#Performance_Considerations