Two questions about the name server address sorting functionality

[Kevin Darcy]

Mark_Andrews at isc.org wrote:

> >
> > "Johan Larsson (EAB)" wrote:
> >
> > > Hi Mark!
> > >
> > > Do you know when BIND 9.3 will be available?
> > >
> > > I also would like to know if the following features are going to be support
> > ed in BIND 9.3:
> > >
> > > True Round Robin (not only random-cyclic as in BIND 9.1.3)?
> >
> > Do you *really* want *true* round-robin? With *true* round-robin, whenever yo
> > u have one
> > address fail, the address which is rotationally "after" that address, gets ha
> > mmered by all of
> > the clients which fail over, e.g. in a ABC/BCA/CAB sequence, "B" gets hammere
> > d with as much
> > as 2x normal traffic if "A" fails.
> >
> > AFAIK, the optimal sort order is "permuted", i.e. all possible permutations p
> > resented with
> > equal frequency (e.g. ABC/BCA/CAB/ACB/BAC/CBA, for a 3-valued RRset), so that
> >  the traffic is
> > as equally spread as possible, even in the face of failures (in the same fail
> > ure scenario,
> > "B" only gets max 1.5x traffic, and, unlike "true" round-robin, this ratio de
> > creases the
> > larger the RRset gets). Unfortunately, "permuted" sort order requires storing
> >  a significant
> > amount of state about each multi-valued RRset (the number of permutations goe
> > s up factorially
> > with the number of records in the RRset), so it tends to be rather resource-i
> > ntensive.
> >
> > Still, if someone is coding "true" round-robin anyway, perhaps they might giv
> > e some
> > consideration to adding "permuted" order as well...
>
>         Well "permuted" is a lot of work and really does not provide any
>         benefit over "random" which is available in BIND 8 and BIND 9.3.

Permuted and "true" round robin are less susceptible to random "spiking".


- Kevin