Load Balancing Question...

[Andras Salamon]

On Tue, Apr 25, 2000 at 06:21:50PM -0400, Kevin Darcy wrote:
> If you reduce the TTL on the RRset then the effect of intermediate caching
> servers is minimized, the downside being that you increase the query traffic and
> just generally be a bad net citizen when you do this. But it may be a perfectly
> acceptable approach on a private net or internet. IBM's mainframe

For the specific case of lower-capacity servers or links (in contrast
with needing to smooth out short-term load fluctuations), how about
using longish TTLs, but different TTLs on each record?  For a weighted
loading in the ratio 2:1:1:2 between four IP addresses,

example.com.  A  86400  10.0.0.1
example.com.  A  43200  10.0.0.2
example.com.  A  43200  10.0.0.3
example.com.  A  86400  10.0.0.4

The intention is that the higher TTL records will stay in the caches
longer.  If I recall correctly, these records will be treated as separate
entities when flushing old cache entries, so the longer TTL records will
indeed stay longer.

I suspect the weights would not reflect the exact ratios, since some
caches would just be asked a few times for the information, and the
RRs would then languish until all expired from that particular cache.
But for very popular domain names the weights should approach the ratios.

Clearly smoothing out short-term fluctuations in server load can't be
done with this approach, and there may be interesting side-effects if a
large ISP cache reaches the point where it starts throwing away the lower
TTL records in its cached RR set while retaining the higher TTL records.

(Surely someone has used this approach before?  And didn't Robert Elz
rant against the evils of inconsistent TTLs a few years back?)

-- Andras Salamon                   andras at dns.net
-- http://www.dns.net/dnsrd/        DNS Resources Directory
--                                  The online DNS reference since 1994