LANs /

State of the LAN: What Ethernet Can't Do

By Kevin Tolly
Network World, 02/19/01

Sometimes the simplest questions are the best. On our last Network World/Tolly State of the LAN tour, a network manager in attendance said the following (and I paraphrase): "OK. Ethernet is great. There is huge hype surrounding it. We know what it is great at, but what are the things it cannot do?" Fair question.

Clearly, Ethernet-at-whatever-speed is not a panacea. There have been competing LAN technologies such as FDDI and token ring that had - at least at some point in time - some demonstrable benefits over the then-current Ethernet alternative. Having been buried by the Ethernet avalanche, any discussion of its benefits is largely academic. Given that the only real alternative to Ethernet is ATM - and only in certain situations - let's explicitly take that context and start with a little history.

During the past decade, Ethernet's architects - the members of the IEEE 802.3 group - have worked diligently at closing the gap between Ethernet and ATM. They've done an impressive job in an array of areas.

In the early to mid-'90s, Ethernet was an order of magnitude slower than ATM with 10M bit/sec Ethernet "competing" with OC-3 155M bit/sec ATM. Of course, raw speed has long since ceased being an issue with Gigabit, and soon 10-Gigabit Ethernet, providing more than ample bandwidth even for demanding metropolitan-area network applications.

The most significant - and successful - encroachment occurred when switched Ethernet devices appeared on the scene. This approach multiplies rather than divides bandwidth by the number of devices on the network and is the essential ingredient for building truly scalable networks.

These changes, along with the price benefits of using copper (because virtually all ATM connections were fiber) and avoidance of ATM's massive complexity, doomed ATM as the "end-to-end everywhere" solution it was long purported to be.

But, to return to our initial question, what Ethernet is not is a circuit-switched system. Yes, one can buy Ethernet switches and ATM switches but Ethernet uses switch packets and ATM uses switched circuits - that makes all the difference in the world.

In traditional Ethernet, there is no visibility beyond the packet once it passes through the port of the switch. Switches don't talk to each other and paths cannot be carved through the network.

So, the "quality of service" offered via Ethernet's 802.1p protocol really isn't that. It is actually a "class of service" mechanism - offering service that is relatively better than others but without any absolute guarantees. The Ethernet offering is not even in the same class as ATM.

Reconvergence and redundancy are other areas where Ethernet pales in comparison with ATM. By the time spanning tree reconverges, another ice age could begin. Again, various efforts within 802.3 - as well as proprietary vendor efforts - are under way or in place to mitigate this problem, but the lack of a "circuit-centric" approach makes the task daunting.

Bulking up connections by using parallel redundant links is possible with Ethernet these days but, although it is ruled by an 802.3 standard, it is still somewhat kludgy and limited by age-old Ethernet architectural issues.

I'd be remiss if I failed to remind you that the 802.3 standard still only recognizes a 1,522-byte frame as the largest chunk of data an Ethernet frame can transport.

This hasn't changed since 10M bit/sec Ethernet. Relative to the transport speed, this is like having the largest frame on 10M bit/sec Ethernet be 1.5 bytes.

Multi-protocol Label Switching, which will allow a circuitlike approach to packet switching, is where we need to look now. It is only by becoming more like a circuit-switched environment that Ethernet will be able to deliver advanced functions.