Beyond metro Ethernet

Why don't we use SDH in the LAN? After all, we use it outside the LAN as the Layer 1 technology to deliver point-to-point services and to support Layer 2 technologies such as Frame Relay for point-to-multipoint networks, and ATM for full mesh networks.

The answer is simple. Ethernet, unlike Frame Relay and ATM, straddles both Layers 1 and 2 of the OSI 7-layer model, so it doesn't need a separate Layer 1 technology.

Also, why haven't we been using Ethernet in MANs and WANs. CIOs, CTOs and IT managers use Ethernet everyday, so why use a different technology outside their LANs?

The usual reasons include the following:

It's not "carrier class" (i.e. it's not expensive and complicated)
It doesn't scale (limited by the 7-hop limit)
No Quality of Service (only needed on bandwidth-restricted legacy networks)
It doesn't have the performance required (we'll see about that oneU).

Without these objections, we would have a ubiquitous network technology that would make terms such as LAN, MAN, WAN redundant since they refer to limitations of the technologies rather than a customer-led desire.

Here we outline options from UK Ethernet Service Providers, e.g. Neos (London metro and UK national), Easynet, Packet Exchange (pan-European and transatlantic), Exponential-e (London metro), and Fibrenet (UK national).

The "fully stacked" model (Fig 1 above) will present Ethernet to the customer. Content is in TCP/IP, clearly Layers 3 and 4 of theOSI model, and needs to be carried across the network. Ethernet can do the job on its own as it's both Layer 1 and 2, but in this implementation it is being carved up into ATM cells. As the ATM is Layer 2, it needs an additional Layer 1 technology to support it — enter SDH.

As this service is SDH-based there are limited bandwidth options available, typically restricted to 50Mbit/s, with a Frame Relay style "burst" mode up to 100Mbit/s. SLAs can't guarantee latency in this heavily overlaid environment so new delay-sensitive applications such as IP Storage can't be supported. ATM cell tax drags down network performance, reflected in the SLA. Services based on this hierarchy have poor availability figures of 99.9%. Due to the cost of the hierarchical stack, service prices are higher than for equivalent SDH bandwidth.

The benefit is that service providers can claim to be in the Ethernet arena.

There's little difference between "fully stacked" and "partially stacked" models (Fig 1 centre). The ATM layer has been omited but the service is still SDH-based. This duplicates the Layer 1 function within Ethernet (known as PHY) and adds unnecessary cost and complexity.

SDH provides few bandwidth options with big steps between each — 2, 34 and 155Mbit/s. SDH-based Ethernet services must reflect this. If a user needs 6Mbit/s bandwidth they must take 3 x 2Mbit/s SDH circuits and concatenate the bandwidth. So, routers still need expensive WAN ports and, to upgrade to 8Mbit/s, a truck roll and 60 days for a further 2Mbit/s circuit.

Such services typically have limited bandwidth options and no latency guarantees. Also, due to the SDH base, they are expensive. Low latency guarantees are important since many applications are delay sensitive and will struggle to work in a high-latency environment.

The "fully delayered" model (Fig 1 right) makes full use of Ethernet's Layer 1 and 2 properties (as in the LAN). But this hasn't happened before due to limitations of "classic" Ethernet, mainly caused by Spanning Tree, which prevents loops by allowing only one physical path through the network. It statistically determines which port each switch/bridge should use to forward through the network. Network convergence times range from fairly lengthy to extremely lengthy, i.e. from tens of seconds to minutes.

Spanning Tree results in wasted bandwidth and poor performance on some routes. As classic Ethernet is a broadcast technology, Ethernet frames containing IP packets can go anywhere, giving low security, unless VLANs are used (limited to just over 4000 VLANs). Also, there's a recommended soft limit of 7 hops across the network otherwise convergence times exceed critical limits and data will be lost.

These limitations can be overcome by using the Martini Draft of MPLS (Multi Protocol Label Switching) with Layer 2 Ethernet switches. An MPLS network consists of Label Edge Routers (LERs) at the edge and Label Switch Routers (LSR) in the core. Layer 2 MPLS adds an MPLS tag and the addresses of the Source LER and Destination LER.

The first MPLS switch encountered is the LER. This uses a pre-configured best path (Label Switch Path) through the network. It controls where the packet goes, enabling security. The LER can support a back-up LSP. If a failure occurs along the main LSP then it sends the data down the back-up LSP with sub-50ms network convergence times.

So, MPLS allows us to:

pick the best path through the network, ensuring low latency, efficient use of bandwidth, and extension beyond the 7-hop limit;
guarantee 10ms latency in a standard SLA;
deliver sub-50ms fail-over times;
deliver security without VLAN's 4000 limit.

We can now deliver a high-performance network using Ethernet, as in the LAN. But development has focused on the metro area whereas users have premises country-wide.

There are two options for nation-wide users:

connect metro networks using SDH, though this is costly and loses all benefits of Ethernet;
use "long-haul" GigE cards.

The latter creates economic high-performance legacy free networks as it enables Ethernet to be transported directly over DWDM rather than SDH. Neos guarantees 10ms latency on its Neosnet national optical Ethernet network at low cost: typically 100Mbit/s for the price of 34Mbit/s SDH and 10Mbit/s for the price of 4–6Mbit/s SDH.

Such an optical Ethernet network can deliver Liquid Bandwith, user-adjustable via the Internet in increments from 1Mbit/s to 1Gbit/s and delivered within 48 hours without having to over-provision bandwidth "just in case".

Last mile delivery options include: