DHCP Failover and Performance

[michoski]

On 2/22/12 11:21 AM, "Chris Buxton" <chris.p.buxton at gmail.com> wrote:
> The performance goal of the failover protocol was to have two servers perform
> nearly as fast as one. That is, the failover protocol itself should impose not
> much more than a 50% penalty on per-box performance.
> 
> In my experience, the performance of a failover pair can range from about 70%
> to about 130% of that of a single box with the same config, with the variation
> based apparently on DDNS configuration. Without DDNS, it appears to usually
> come in around 130%.

Indeed!  Beside the logging advice already posed (which, in fairness,
affects many/most applications similarly...not just DHCP -- try running
OpenLDAP with log levels cranked up at a busy site and watch it burn),
separation of core functions is a big part of scalability.  In general, I've
always heard "best practice" is to fully separate (as it what disk spindles
it hits, not just more virtual instances on the same disk!) DHCP and DNS
(DDNS as well as any secondaries which do zone transfers).

> When you say "dramatic effect", what exactly do you mean? Can you be more
> specific?

I've noticed certain commercial solutions (which you'll pay dearly for) put
a lot of effort into things like "avalanche protection" and can often
support ridiculous numbers of managed IPs (>1 million) per two-node
"cluster" with floods of 50-70k renewals being quite common.  I'm not versed
enough to know if ISC DHCP has similar functionality, but "dramatic effect"
certainly sounds like the "lease renewal avalanches" I've read about.

-- 
Whenever you find that you are on the side of the majority, it is time
to reform.
        -- Mark Twain