Description of problem:

On deployments that have large neutron environment requirements, customers experience timeouts and other issues related to arp tables being exhausted due to low default Linux kernel caching values.

These problems are trivially solved with kernel tuning, but should be avoided entirely by making such tuning default during deployment of neutron networking in the overcloud.

Version-Release number of selected component (if applicable):


How reproducible:

100%

Steps to Reproduce:
1. Deploy overcloud in large network environment

Actual results:

Ping timeouts, other general routing failures.

Expected results:

Normal network functionality

Additional info:

Packet loss in such an environment, pre-tuned, resembles the following:

64 bytes from 192.168.10.2: icmp_seq=150 ttl=64 time=0.054 ms
64 bytes from 192.168.10.2: icmp_seq=151 ttl=64 time=0.066 ms
64 bytes from 192.168.10.2: icmp_seq=152 ttl=64 time=0.051 ms
ping: sendmsg: Invalid argument
ping: sendmsg: Invalid argument
ping: sendmsg: Invalid argument
ping: sendmsg: Invalid argument
ping: sendmsg: Invalid argument
64 bytes from 192.168.10.2: icmp_seq=158 ttl=64 time=0.972 ms
64 bytes from 192.168.10.2: icmp_seq=159 ttl=64 time=0.035 ms

This is resolved with the following tuning:

Add the following lines to /etc/sysctl.conf:

net.ipv4.neigh.default.gc_thresh1=1024
net.ipv4.neigh.default.gc_thresh2=2048
net.ipv4.neigh.default.gc_thresh3=4096

Then run:

sysctl -p

Immediate and positive effect will be observed.

The purpose of this RFE / bug request is to see about having these tuning values added; higher defaults would mean customers never encounter these errors.