In my early days in this industry, there was a topology battle between
Ethernet and Token Ring. Token Ring was just about to release an
amazingly fast 16MB/s version. It was interesting to watch IBM
representatives, still in suits and ties then, draw a circle on
the board and make analogies to Token Ring and the railroad. They were
making the case that Token Ring was going be the infrastructure of
choice and that Ethernet's time had come.
In one such session, I made the comment that while the railroad is/was
known for precision, an airplane can start late and land late but still beat
a train to its destination. Speed often wins over precision and as we
all know Ethernet won that battle, handily. Along the way there have
been many attempts to dethrone Ethernet and they all just keep failing.
The closest battle was using Ethernet as transport for storage protocols
vs. SCSI or Fibre Channel. While I expect Fibre Channel in its native
form to continue on for some time, Ethernet will eventually carry the
majority of storage traffic.
The only real stumbling block for large scale FCoE deployments is the
complications of Spanning Tree Protocol (STP). As we discuss in our
recent article, "What is TRILL's Role in FCoE Storage Networks?", Transparent Interconnection of Lots of Links (TRILL),
should fix many of the issues that storage and large networks in general
have with STP. These problems are being exacerbated by the convergence
of storage onto Ethernet and the increasing workload of server
virtualization. TRILL will provide us with multi-path layer 2 networking
allowing for use of the idle network bandwidth that STP essentially
blocks.
Of course, prior to FCoE, we have already seen the growing use of NAS as
more than just a file-sharing protocol. We have seen it used for
databases and virtual machine images. iSCSI of course still continues to
move along providing block access to storage across IP. Finally, new
uses of Ethernet like ATA of Ethernet (AoE) are bringing high-performance
to the protocol without as much overhead. In short, the storage protocol
choices on Ethernet are astounding.
I/O
Virtualization (IOV) technologies and cluster interconnects that leverage
high-speed Ethernet now threaten the PCI/e and
Infiniband architectures. The advantage that the Ethernet systems have is built-in connectivity and scale with a readily available chip
set. In theory, versions of this interconnect technology could be built
right onto server motherboards. Imagine a server with built-in
IOV capability or a special server class designed for clustering with
the server interconnect already built in. Will Ethernet win in these emerging use cases? There is no guarantee, but
given Ethernet's track record, I don't think I will be betting against
it.
