re: Juniper's TX Waits Its TurnWith 50 Mbit/s to each home as discussed elsewhere on this web site, http://www.lightreading.com/do..., and with about 20,000 subscribers for a typical central office, you got 1 Tbit/s of bandwidth needed in an edge router! Granted, you will do some sort of statistical multiplexing, etc., but the point here is that bandwidths of the order of Tbit/s could easily become routine in the not too distant future.
re: Juniper's TX Waits Its TurnI don't think you can dismiss stat muxing quite so quickly, krbabu. Here in the UK the typical broadband contention is 50:1 for home users and 20:1 for a business service.
You're describing a 50Mbit/s-per-end-point service that I assume would continue to be over-subscribed at these kinds of levels.
So your aggregation router (B-RAS or IP Services Switch, whatever it becomes by that time) actually only needs a capacity of 50Gbit/s (that's at 20:1).
And if we're still talking asymmetric service then that normally means one third of the bandwidth upstream compared to downstream. So a 50Gbit/s full duplex box might not even be breathing hard.
And remember you're talking a CO box, and so the SP could always distribute this load over several boxes, for resilience or load reasons.
So from these simplistic numbers we can already see that things could easily flip the other way if...
1. Contention rates drop. Quite likely if SPs expect fancy VoD services to work properly. But if you're whacking up the bandwidth to 50Mbit/s then contention rates may stay high. I suspect they will need to stay high in order to enable "IP economics", with premium customer flows being prioritised through the contention using DiffServ.
2. Asymmetry assumptions change. You're talking about a 50Mbit/s service, so that implies about 15Mbit/s upstream. I can't see me generating more than this amount of bandwidth from home, or even a home office - although it might finally mean decent video conferencing is possible. Real business offices will use SDSL, of course, and that's a different story.
re: Juniper's TX Waits Its TurnDoes anyone need a big router? At what point do you just strip the headers out of IP traffic and put giant circuit-like streams on optical switches? Seems as though a logical setup is lots of electrical/smart switching in MSAs, then rail-like optical transport between them. With truly large streams and low prices, are giant routers with $1M blades affordable?
What can GigBE teach us here? SCMR, where are you?
First, IP traffic (unlike the PSTN) does not seem to generate a great deal of geographic locality. That provides added stress on the cores.
Second, while the need is not for immediate usage today, there is a clear future need. Carriers have been paying for many generations of forklift replacement of core routers. They dislike this both from a cost perspective as well as an operational perspective. They would very much like to have a system that can grow incrementally, without the forklift. While the immediate market is for single units, the eventual growth of these installations is quite clear to all. The carriers are quite adamant that they want to see working product before buying into any scalability story, so the vendors do need to pony up with working scalable architectures before they sell the tiniest of installs.
You're absolutely right, as the scale of the network increases, it pays to increase the density of the topology and avoid intermediate switching steps. However, that doesn't mean that you can get rid of them completely. At some level if you want to aggregate traffic to take advantage of the stat muxing, then you also have to deaggragate that same traffic.
Thus, you have something like a metro area that needs to eventually absorb something like terabytes of bandwidth and distribute it.
re: Juniper's TX Waits Its Turn> so the vendors do need to pony up with working > scalable architectures before they sell the > tiniest of installs
True, multichassis capability might boost sales just by itself, just like the halo cars do for the auto manufacturers.
But, per definition, multichassis is only effective in scenario where traffic growth outpaces traditional switching architectures. As soon as the latest silicon catches up, multichassis instantly loses to a faster and more compact single-box design. It sounds extremely unlikely that someone will ignore forklift upgrade option in order to invest into dozen or more bulky, unholy expensive and unreliable VCSEL interconnected racks. So, unless the volume of traffic will start outrunning the Moore's Law again, multichassis behemoths might remain as rare and practical as concept cars, being so wrong to everyone but few.
re: Juniper's TX Waits Its Turn First, the volume of traffic continues to be about 100% per year, which is well above Moore's law.
Second, Moore's law doesn't apply to bandwidth. In particular, single box switch architectures are limited by signal integrity issues and serial link speeds, not by the speed or density of the transistors.
Third, yes, multichassis may not be the ultimate "correct" answer in engineering terms, but you need to recall that the customers are looking for investment protection, not optimal technology. The forklift upgrade plan has them replacing their systems every 18 months, which is considerably faster than their depreciation cycle. You need to integrate this out over a 10 to 15 year horizon to see the implications. Also, don't forget that technological advances in the individual chassis DO get applied, the only required constant is the interconnect architecture and the ability to perform a rolling upgrade.
re: Juniper's TX Waits Its Turn> Also, don't forget that technological advances > in the individual chassis DO get applied, the > only required constant is the interconnect > architecture and the ability to perform a > rolling upgrade
This is precisely why I do not see the need for anything beyond low-volume 4x or 8x multichassis systems for the high end of the market. Interconnects are not free of the signal integrity issues found in traditional systems, they just use the brute force approach to fight them (i.e. surface lasers vs. copper traces). On top of that, interconnected shelves always need more supplemental hardware (supervisor modules, power, sheet metal, etc) when compared to a single box, while a traditional router may still avoid forklifts by a series of technology upgrades. Furthemore, if the real-world demand will outrun the state-of-the-art tech so much that multichassis interconnects will become common, the pain of TCO may soon become high enough to generate the need for another breakthrough startup.
re: Juniper's TX Waits Its Turn Well, what can I say? The industry does not have a good track record of delivering technology upgrades without forklifts. In fact, it doesn't even have a good record on technology upgrades even WITH forklifts.
I suspect that this perception and the mission critical nature of the core router is sufficient to preclude any argument that requires that the vendor be trusted to deliver.
Assuming you have the volume/size to actually sustain multiple terabits, aren't you routing a good majority of that closer to the edge anyways today ?
= K