In 1994 ATM was the 'acronym of the year'. We were told that by the Millennium, all telecoms services would be ATM based, and all software applications would communicate using ATM APIs, to exploit its uniquely rich set of features for quality of service (QOS).
Yet today, all the major application vendors are feverishly writing or converting software and applications to use the TCP/IP protocol suite. ATM's real problem is that it has not delivered in time, mainly due to the slow emergence of the necessary standards. It lost ground very early as the end-to-end technology of choice. This has resulted in the introduction of considerable complexity as manufacturers have struggled to make ATM a viable technology for niche solutions.
Current thinking amongst the market leaders is that ATM's cell-switching is as good as dead, killed by simple packet technology used end to end, namely TCP/IP. This is confirmed by the Conspectus survey - see previous article- which found that 84% of respondents were either using or evaluating TCP/IP as the network protocol of choice.
The latest high-speed packet switching and transmission techniques are consigning cell and circuit-switching to the dustbin of history along with the nightmare of supporting compromises and complexities (e.g. MPOA, MPLS).
We are moving to a position where a ubiquitous TCP/IP-based infrastructure will deliver the next generation of voice, video, data and true multimedia services. There will be no cell-switching (or circuit-switching) and ATM will be just a distant memory of yet another failed standard.
The only parameter that will remain relevant to measuring the QOS of multimedia services will be end-to-end delay. Until recently, transmission and switching technologies have been significant contributors to delay - and this has historically been overcome by allocation of 'bandwidth paths' through end-to-end circuit-switching.
The circuit-switching camp cling to the premise that the way to minimise delay is through the pre-allocation of bandwidth, with ATM providing the virtual circuits (albeit through a cell-switched fabric). The packet-switching (IP) world disputes the requirement for per circuit bandwidth allocation so long as the individual communication streams are small in comparison to the size of the transmission paths.
To a large extent, this is a question of scale. Recent developments mean that:
- Wave Division Multiplexing makes it possible to carry the entire planet's combined voice traffic on a 16 core fibre cable
- Commodity routers can forward almost 1,000,000 IP packets per second (equivalent to more than 10,000 voice calls).
Of course, even if transmission and routing are no longer a significant cause of delay, queuing caused by congestion can still occur. In the connectionless IP world there are two possible approaches to managing bottlenecks:
- Build networks where the queue delays are reduced by increasing the size of the output channels in relation to the inputs.
When bandwidth was the principal network cost, network designers rarely considered the simple approach of making the network bigger. However this is changing and we can already see how technology has shifted this basic premise. Today, in the developed world at least, we never hear the network busy tone on the PSTN, which really means that the voice networks are already over-provisioned. The same technique is equally applicable for packet-switching. Better still, in a connection-less IP network there is not even the concept of a busy tone. In general, even if there is some congestion a user would rather have a reduced quality of service as opposed to no service at all.
- Prioritise traffic to expedite delay sensitive communications by managing the queues.
Several end-to-end streams based QOS solutions (comparable to the ATM virtual circuit solutions) have been developed for IP. The best known is RSVP, but all have proved difficult to implement. The latest proposals for QOS management for IP are much less complicated. The schemes are based on a simple system of defining a small number of network-wide frame identifiers on which to base forwarding priorities. QOS will be managed on a hop-by-hop basis, on the assumption that in a generally over- provisioned network, prioritising traffic at each hop on the basis of a high/medium/low sensitive to delay will be sufficient.
In summary, the complex, connection-oriented infrastructures of today will soon be replaced with simple, connectionless technology that will rely upon high-speed transmission and lightning quick routing. The equipment deployed will be inexpensive and simple to manage.
Manufacturers of complex switches beware: the tide has turned away from elaborate traffic management schemes ('small and smart') in favour of over-provisioning ('big and dumb'). As we overheard at an exhibition the other day: 'Circuit switches are not dead yet - but only a fool would invest in new ones'.
