A case for using hybrid optical switching in long-haul networks

There is a fundamental difference between using transparent elements at line sites — degree two junctions with only two fibre directions — and using them at switch sites that have a degree greater than two with three or more fibre directions. If we are optimizing total network costs, we need to consider the impact that the switches will have on the line system costs.

The first such impact is on the reach requirement of the line systems. We might imagine that this is a simple matter of understanding the distribution of demand distances, and then using the appropriate ultra-long reach transmission technology so that most of those demands can be met without using those evil regenerators.

The common argument goes something like this: "80% of the demands in North America are 4,500 km or less; we have ULR capable of 4,500 km: therefore we can build the network largely without regenerators and obviously it will be much cheaper." This is slightly disingenuous. The extra reach doesn't come free. ULR can be achieved by using better transponders than in LR systems, or by using distributed Raman amplifiers in conjunction with the normal erbium doped fiber amplifiers (EDFAs), or by using both techniques together. The point is that both techniques increase costs.

How much more expensive is the extra reach? ULR transponders capable of 200% of the reach of LR transponders cost roughly 50% more than LR transponders. And, equipping a line system with Raman preamps plus EDFAs will roughly double the amplifier cost of that line. In a lightly loaded (few wavelengths) ULR line system, the amplifier costs will dominate (many amplifier sites, few wavelengths). In a heavily loaded ULR line system (many wavelengths) the transponder costs will dominate, especially where the transparency distance requirement is less than the maximum system reach (fewer line amplifier sites.)

A further subtlety is the granularity of reach. Assuming a constant cost/bit/sec/km, compare a transmission system with 1000 km reach against one with 4500 km reach. Given a spread of demand distances from say 250 km to 6,000 km, with for argument's sake a normal distribution, it's easy to see that the 4500 km reach system would only be cheapest for demands between 4000 and 4500 km, and possibly as one segment of demands above 4500 km. All others would be cheaper using the appropriate number of 1000 km segments with repeaters.

The bottom line: savings that might be expected by reducing the number of transponders in a transparent network are offset by the higher cost of the ULR transponders and line systems required. We save less than might be expected just by counting transponders.

Switching is used to make more efficient use of the transmission systems, in three ways: sharing working bandwidth when the demands change, sharing protection bandwidth and grooming demands to more efficiently pack wavelengths.

Generally, protection path lengths are longer than the shortest path from source to destination, because the shortest path is used for the working bandwidth. To be effective against backhoe fade, the protection path must be different from the working path. In some cases, the protection path may be much longer than the working path, depending on the fibre topology of the network in question. And, if we are doing the protection switching transparently, then the transmission path from laser to detector will change on the protection switching event. If we are using some of the more efficient protection schemes — mesh protection, for example — then the protected path length may not be easily predictable, it will depend on instantaneous network load and on the location of the failure or failures.

There are two corollaries. The first is that this effectively forces all of the links in the network to be designed for worst-case path length. The best available ULR technology must be used, even on the shorter links. This drives network costs up — and not because of the switching costs directly, but because of the impact that the switching technology has on the line systems design.

Regenerating the signal

The second corollary is that we will need to add transponders to the network for selective regeneration. Since we can't achieve infinite reach, even for infinite money, there will be cases where the required reach exceeds the available reach, following a protection switching event or even a normal demand routing event. In this case, we must regenerate the signal, and the only cost-effective way to do that is with a transponder.

It's a complicated design problem to predict where the selective regenerators might be needed, so that transponders can be pre-provisioned at those nodes. And, the result is that we have to populate the network with a greater or lesser number of transponders which will sit idle until required to complete a demand route.

Transparently switched networks cannot groom. This doesn't matter if the services provided by the network are simple bit-streams at the modulation rate of the wavelengths (10G or 40G), and if all the services and transponders are the same across the network. But it does matter if the services offered include any OC48 or OC48c services, since it's no longer cost-effective to modulate wavelengths at 2.5G, and it's inefficient to burn a 10G or 40G wavelength for a 2.5G service. Or if the services are 10G, but the cheapest transmission method happens to be 40G.

The secondary argument concerns edge versus core grooming. Some argue that all of these efficiencies can be realized by performing the grooming as an edge function, before the core transport network sees the traffic. The real answer is "it depends." If the number of demands is large compared with the number of nodes (sources and destinations), then grooming won't make such a big difference. For each destination, there will be many wavelengths full of traffic from each source, and zero or one partly filled wavelength. Core grooming might allow that last wavelength to be groomed with other partly filled wavelengths at intermediate nodes, but the efficiency gain is small.

It's important to note that all of the design challenges discussed will vanish if a single architectural decision is made: to use grooming OEO switches at the degree n fibre junctions where there are more than two fibre directions meeting. The whole "arbitrary optical path" problem vanishes, and the transmission link design problem reduces to a fixed path, with known length and dispersion.

Noting that those design challenges don't apply to links with only degree two nodes, and considering that there may be more of these nodes than higher degree nodes in a complete North American transport network, we might well ask if there isn't some cost saving to be made by using transparent elements at the degree two nodes. It's a good question, and the answer is that considerable cost can be removed by using OADM network elements at these locations, if the percentage of add/drop traffic is less than maybe 70% of the total traffic through that node. There are still challenges, but less severe ones.

To be efficient in a dynamic demand environment, the OADMs need to be reconfigurable. OADMS add a transmission penalty in the through path: if there are several OADMs in a link, ULR transmission technology may be needed where LR would have been sufficient. Even solving these challenges, OADMs still prove in for many degree two nodes in the North American networks that we have designed.

Couldn't we combine a degree two OADM and a grooming switch at degree three or higher fibre junctions? By constraining the transparent switching to just pass/drop/add, we still avoid all of the design challenges that degree three or higher transparent nodes introduce. This does turn out to be useful in reducing costs in some cases. It's worth noting that in the case of Altamar's Titanium product, the OADM is integrated with a scalable grooming switch, even when used in a degree two node. This allows more efficient use of transponders for the add/drop traffic, and of course grooming as well.

Many believe that the high costs of transparent networks are simply a function of the maturity of the optical components used, and that over time, transparent networks will become cheaper than hybrid networks. This turns out not to be the case. First, it's not the costs of the switches that matter: almost all of the network costs are in the line systems, and most of the line system costs are in the transponders.

Transponder for life

The first future cost problem that transparent networks face concerns the rapid evolution of transponder technology. As we have seen, a non-blocking transparent network can only use one transponder type. This means that once chosen, the same transponders must be used for the life of the network. We can't take advantage of say 40G transponders if we have built the network on 10G. We can't even use new modulation formats on 10G.

In contrast, the opaque or hybrid networks can use new technology on a link by link basis, and with some constraints, even on an existing link. Since there are grooming OEO switches at the ends of the links, transmission technology used on one link has no effect on the technology choices for the next link, or any other link. We are always free to optimize the transmission technology for lowest cost for each link, at any time.

Therefore, over time, the transparent network built today will be at an increasing cost disadvantage to the hybrid or opaque networks. But what if we just look ahead a few years, and compare the costs of building a new network then, with transparent or hybrid or opaque switches? This brings us to a slightly different problem. Transponder technology is evolving very fast, and the cost of building a 10G LR transponder has dropped by a factor of ten over the last few years. There are developments in process which will give us another factor of ten over the next few years, with another tenfold decrease conceivable after that. So, the "$100 transponder" is at least conceivable.

But here's the key fact: the technologies which allow this kind of radical cost reduction do not allow ULR performance. And the economics of ULR are balanced on a knife edge. While we can get twice the reach for 50% premium (considering transponder costs only, not the extra line costs), ULR proves in for at least a part of the network. But if the cost ratio changes slightly — say twice the reach for 80% premium — then there won't be any ULR used. It would always be cheaper to regenerate. And all the signs are that this cost ratio is diverging, not converging. The lower performance LR transponders are getting cheaper faster than the higher performance ULR transponders.

And since transparent networks depend on ULR transmission for their economics, they will be at an increasing disadvantage.

This has nothing to do with the cost or scalability or reliability of MEMS switch fabrics. Even if transparent switches were free, infinitely scalable and infinitely reliable; it would still be cheaper not to use them.

The full version of this article was presented at the 2002 Communications Design Conference.