I'm using policy routing to load-balance outbound traffic across two Internet links.

Everything works fine, but it caused a problem when I upgraded ntp from 4.2.4 to 4.2.6: the later version resets the peer state when the source IP address changes, and I find that the policy routing load-balancing causes the outbound UDP packets to change source address around every ten minutes. This means that ntp keeps losing sync.

I think this is an ntp issue, but I've discussed it on the comp.protocols.time.ntp usenet group and there doesn't seem to be an application-level solution. Ref subject "ntp server with two default routes misbehaving after upgrade" if you're interested.

I've worked around the problem for now by replacing the uk.pool.ntp.org servers in my ntp config with fixed IP addresses and using host routes to send all traffic to these hosts via one route. That's ok as a short-term measure, but isn't a good long-term fix.

I was wondering if there's a way to tune the route cache to make traffic for a given destination stick with one route. I thought routes were cached for 300 seconds, but I find that even reducing ntp maxpoll below this time doesn't cause the routes to stick, so there seems to be more to it than that.

I can't find much in the way of documentation for the route cache, so I wondered if anyone has any pointers, or ideas on how the route cache can be tuned to achieve what I want.

Can't you add a policy that sends all traffic for dest port 123 on one interface?

I did think of that, but it's not ideal as I lose the resilience of the multiple links for NTP, and I suspect it would prevent the external routers from ntp-syncing back to the host.