|
|
|
The MetaRouter Concept
Topology: Our topology is based on
group theory.
It reduces the number of rows needed to achieve complete
connectivity. A Banyan or Butterfly network requires log 2 N
rows to achieve complete connectivity. Our topology achieves complete
connectivity in log X N rows, where x is half of the fanout of
a switching module. This difference in topology
reduces the hardware
needed to build a multi-stage switching network by more
than a factor of 2.
We call this topology a Galois Network, named after the French mathematician,
Évariste Galois. This was easier than constantly referring
to the network as a "blocking compensated cyclic group networks."
Routing Algorithm: Our
destination-based routing
algorithm is similar to that used by the Internet. Each
switching module uses its own table to determine how a packet is routed.
Our routing algorithm requires more hardware than the routing algorithm
used by a Banyan or Butterfly network. It also requires global network
information to update the tables. However, our routing algorithm can
handle a wider range of topologies; exploit external network redundancy;
and tolerate more faults in the network.
Switching Module: Our
switching module consists of buffers, a crossbar, and decision logic
which analyzes the packet header, looks up the destination address in a
table, and schedules the crossbar.
We noticed
that the functionality of our switching module was similar to that of an
off-the-shelf IP router. We then realized that we could avoid having to develop
custom hardware via a judicious choice of module interfaces and
algorithms. This can be accomplished by:
using an IP bit stream
format between the modules instead of a proprietary format; using PPP (point-to-point
protocol) for flow control instead of a proprietary hardware handshake;
and using IP routing between
modules instead of a proprietary routing algorithm.
In perspective, any terabit router can be deconstructed
into "Cisco" equivalents. This follows from the fact that they are all based on
multi-stage switching networks. The only real differences are in the
topology, the routing algorithm, and the packaging. In retrospect, it is
difficult to construct a proprietary hardware position on top of an open
standard such as TCP/IP. (IBM discovered this with their PC.)
|
|
