[nsp] slb predictor

[atticus at satanic.org]

> Are there any important differences between using roundrobin and
> leastconns as slb predictors

Intuitively, leastconns is the "right" approach, though it can be
unpredictable, given its dynamic nature.. Unless you're trying to manage
unbalanced server platforms or very unbalanced request types, I'd avoid
it.

The one caveat I've stumbled across w/ round-robin (and was unable to
convince the TAC of its bugginess) is the default weight. The way its
implemented, each real will get N (where N is weight) consecutive
connections before the next server is selected. Unfortunately, default
weight is 8, so in a "plain" round-robin setup, each real gets connection
in sets of 8. Not a big deal if you're dealing w/ high volume, but it is
if each connection is relatively expensive.