There was a time, not so long ago, when asynchronous transfer mode (ATM) was cited as the networking technology that would unite Wans with Lans. The Holy Grail of that era was a common infrastructure inside and outside the enterprise. However, the marketing machine that has been relentlessly driving forward the adoption of gigabit ethernet in the enterprise - with varied success - has forced network managers to question whether ATM has a future at all.
Peter Boland, vice president of product strategy at Nortel Networks, said: "It is now very hard to fill the pipe and you are never going to run into real quality of service [QoS] issues there. This means that ATM's strengths are not going to be seen inside the enterprise, and it is too complex and costly a technology to compete in a space where it's main strengths don't count."
This has left ATM as a Wan technology, but even in its heartland a significant threat is looming.
"Using DWDM [dense wave division multiplexing] we can now put 160 wavelengths on a singe fibre, to give us 1.6 terabytes of data a second. Most telecoms network operators today are designing their systems around 2Gb, which means that a switch like this gives them 1000 times more bandwidth than they need," says Boland.
"Despite this our factory is on triple shifts - what this means is that people are starting to think very differently about how they make and provision telecoms networks."
In his view, bandwidth will still need to be quality-managed in the medium term, which gives ATM a solid enough present position. But as IP management capabilities evolve with multi-protocol label switching (MPLS), giving IP traffic the ability to carry priority signals, people will start to solve bandwidth management problems with brute force and a little help from MPLS.
Not surprisingly, Tim Foreshaw, director of project marketing at ATM specialist Newbridge Networks - now part of Alcatel - strongly disagrees with Nortel's view of the long-term future of ATM. "We see it as a core technology that will continue moving forward," he says.
For Newbridge, the key argument is QoS in converged networks. MPLS is all very well, Foreshaw argues, but while you can tag a packet with a priority header, you can't do a thing about its length with MPLS.
"With MPLS, you are appending a label to an IP frame and this label steers it through the network. Each individual node does all the Layer 2 and Layer 3 switching," says Foreshaw. "This is good network engineering, but it is far from the whole story."
If IP packets are carried over a routerless network using MPLS, traffic can be directed down a given path, but the variable length frames can't be compensated for, and these will play merry hell with QoS delivery.
Running IP in native mode means you are stuck with variable length frames entering a connectionless, non-deterministic network. Therefore, when things get hairy, users can kiss QoS goodnight. "Frame length could be anything from a few bits to thousands of bytes. You have none of the control over it that an ATM network can offer," says Foreshaw.
By contrast, when IP traffic is carried over ATM, the packet is segmented into determined cells, which gives an even length, all the richness of ATM control data and full QoS.
Moreover, Foreshaw argues that carriers have to support a range of legacy systems. Most of them are not free to act as if Frame Relay didn't exist, for example. He said the best way of handling legacy traffic is with IP over ATM, which is why Newbridge has built IP capabilities and MPLS into the Newbridge 36170 ATM switch.
Raj Alur, director of strategic marketing at Lucent Technologies, takes a dualist view. "It would be ludicrous to pretend that ATM will last forever, but right now it has a very important function to perform. Carriers like it and they are buying it in boatloads today from us and from Newbridge," he says.
He points out that with time-sensitive communications such as point-to-point voice you have to guarantee that no other traffic flows will disrupt the stream or grab precedence. If you go beyond 100 millisecond latencies with voice, the quality crashes. Currently, ATM at the core is the only way of ensuring that no matter how bursty traffic becomes, you can deliver this kind of standard.
He suggests that a look backwards shows why things have developed as they have. A few years ago, when the internet was just starting to boom, big ISPs kept their backbone services simple. However, as the number of ISPs wanting to peer with the existing players grew and grew, the router tables got really messy.
People wanted to add a second hierarchy of routers, giving the net a distinction between core and edge routers. Even so, core routers became bottlenecked arbitrating paths for a growing mass of incoming traffic.
It was time to rethink.
Talk about flattening the network gained currency, and ATM switches came into their own as the ideal way of achieving this. Vendors designed high-speed switches to replace core routers and throughput went up by an order of magnitude. At the cell layer, ATM switches comprehensively outperformed routers.
On top of this companies got real benefits by connecting routers to switches and letting switches set up virtual connections with each other. Where it had been impossible to physically connect all routers, it now became possible to have routers pointed at switches that could handle vast numbers of virtual point-to-point connections through ATM.
There may be trouble ahead
Once the internet started heading for a routed edge and a switched core, router manufacturers spotted trouble ahead. ATM vendors looked like getting too large a slice of the cake. So Cisco, not entirely coincidentally, developed the idea of tag switching, which gave the router the ability to do some of the traffic engineering that ATM switch vendors were planning to handle.
It did not take a lot to convince the market that if you could get switch and router functionality out of the same box, you would make significant cost savings. "The problem was that while we at Lucent were working on getting router functionality into the switches - as were others, such as Cascade - at the same time Cisco was working in the opposite direction," says Alur.
"Cisco submitted tag switching to the standards committee and Lucent and Cascade came up with the idea of protocol switching, using Label Distribution Protocol. This is the mechanism that allows the switch to set up explicit route paths. The whole thing got mixed together and out of it came MPLS."
Like Newbridge, Lucent now builds MPLS-capable switches that can participate in an MPLS cloud, but which run IP over ATM. It looks like an MPLS switch but, down at the guts, it is still ATM. Alur says this means that ATM has a decent future ahead.
Another major trend impacting the industry, says Alur, is the emergence of what he calls "an access network".
"We have all kinds of things going on in the access space. We have cable television, 3G wireless and the unbundling of the local loop for DSL services.
"What the carrier now needs to be able to do is to take in feeds and connections from the customer environment, irrespective of whether these are T1, broadband, fixed wire or wireless."
Alur has to aggregate them and peel them off into different service types at a central office and then feed back all the relevant information to a service provider. Again, ATM's inherent strength in defining fine-grained traffic types makes it ideal as the mechanism that will enable both multi-type traffic aggregation and the peeling off of distinct streams. However, not everyone agrees with Alur and Lucent's view.
Alan Taylor, consulting engineer for Juniper Networks, believes that ATM is already becoming redundant. "The Lucent story about ATM switches at the core of the internet is around two years out of date. It belongs to an era where ATM switches went faster than routers. Now our M40 router goes about 10 times faster than the best ATM switch," he says.
"If you have IP forwarding nodes that are non-deterministic in their forwarding characteristics, which is how routers used to be, then you need something to give you control in a non-deterministic world. However, MPLS and our current generation of routers do the job, so who needs ATM?"
The point about a router such as the M40 is that it is a line rate system, says Taylor. "You can always fill the pipe, whatever the conditions. And with MPLS, you are label swapping at each individual switch. The hops are all decided at the head end router and you create a determined route across the network to the destination from the head end router."
The whole point about ATM cores and ATM virtual circuits was to even out traffic congestion around trunks. However, once ATM switches stopped being able to scale to meet the hugely increased traffic flows on the internet, they themselves became a bottleneck, and this lack of scalability became a new driver to dump ATM.
Not just sour grapes
In case this sounds like a router manufacturer having a swipe at a switch manufacturer, Taylor points out that ATM has a real bottleneck at the segmentation and reassembly (Sar) level. This is the function that occurs when the switch hands off to the packet core.
ATM switches, he says, cannot run a Sar port faster than OC12 (622Mbps).
This becomes a real problem when large ISPs or telcos need to run OC48 (2.4Gbps) and higher connections. "If a router needs to connect to the core at 2.4Gbps, then ATM ceases to be an option," says Taylor. "To avoid the Sar bottleneck, you need an IP packet core.
"Our M160 router, which is just shipping, can run eight 10Gbps ports full duplex or 160Gbps."
According to Taylor, similar arguments apply at the optical layer. Optical networking is going to become a very efficient way of doing transport, and people are starting to want to run MPLS and IP direct over DWDM, without first layering it over ATM.
"Our challenge is to keep up with the packet translation coming off the optics. We can now drive a 10Gbps stream, and that is fast enough right now to make ATM redundant," he says.
Whether this argument will cut much ice in the market remains to be seen.
Ovum telecoms consultant John Matthew argues that the majority of telecoms carriers are still wedded to the idea of keeping ATM in the core. They know and understand its management capabilities, and Sar bottlenecks are not yet too much of an issue. However, he predicts that technology will move on, and that ultimately ATM could be left behind.
Geoengineering on the sea floor near glaciers would form a new ice shelf to prevent melting
Alterations in capillary blood flow can be caused by body position change
Curiosity rover is in 'normal mode' but not transmitting scientific data back to base
NatWest outage comes a day after Barclays' IT systems shut out customers and staff