We run a moderately busy (~200qps) kea-1.6.2. hot-standby pair of servers. Clients are handed dynamic addresses from multiple large dedicated pools serviced via Cisco DHCP proxy, i.e. the Cisco router "spoofs" the servers functionality by rewriting GIADDR and source IP in the server's reply packets. Leases are in memfile.
We've experienced service impairment/outages a couple times whose symptoms are described in kea's congestion handling documentation, although we haven't gone deep enough into tcpdumps to make a hard correlation. Implementation of congestion handling seems to have alleviated the problem.
When problems were occurring, CPU/Memory showed nowhere near capacity, leading me to think that it's possible we could simply do some system tuning to address this rather than the congestion handling.
What I'm looking for here are some general recommendations in terms of best practice, e.g:
- Can we substitute tuning of appropriate system buffers for the ring buffer and if so how?
- If the ring buffer is the best solution, what is the best way to appropriately monitor/size it?
- Any other recommendations?
ISC funds the development of this software with paid support subscriptions. Contact us at https://www.isc.org/contact/ for more information.