ReplEx in itself is not a routing protocol.
Rather, it operates on top of an underlying routing infrastructure of
arbitrary nature, e. g., OSPF, IS-IS, MPLS, static routes, any
combination thereof, ... This makes ReplEx
very flexible to apply.
Normally, if a route can be reached via multiple equal-cost paths,
then traffic is equally distributed between these routes, e. g.,
½:½ or ⅓:⅓:⅓. In contrast, ReplEx adjusts the traffic
splits between multipath routes to arbitrary values, e. g., ⅓:⅔ or 4%:17%:79%.
This means that the underlying routing protocol(s) have to provide
multipath routes — the more destinations are
reachable via multipaths, and the more multipath routes are available
to reach a given destination, the better it is for ReplEx.
ReplEx consists of two parts: The part that
adjusts the traffic splits, and a distance-vector like protocol to
exchange information on network traffic conditions.
ReplEx has been extensively evaluated in
realistic network simulations. These simulations use realistic TCP
workload traffic from a BSD-style implementation (whose on-off sources
follow heavy-tailed processes and thus are hard to handle, because
they generate very bursty traffic). The topologies comprised simple
artificial scenarios to investigate specific behavioural aspects, as
well as more complicated backbone topologies
from Rocketfuel
maps and
the TOTEM
project.
The simulations confirmed the theoretic claims (i. e., quick
convergence without oscillation) even when being confronted with the
congestion control feedback loops of a large number of TCP clients,
and showed furthermore
that ReplEx attains performance improvements
that outrival those that are achieved by traditional static IGP link
weight traffic engineering.
An extensive evaluation is provided in [2], others are given in
[1] and [3].
One of these strategies is the (α,β)-exploration-replication strategy. It can be proven that this strategy quickly converges without oscillations to a stable state [4].
The ReplEx algorithm is an adaptation of the theoretical (α,β)-exploration-replication strategy to real-world communication networks (hence the name, from exploration and replication). An extensive overview on the theoretical aspects is provided in [3].
