Slashdot Mirror
Multiterabit Switching, No Moving Parts
npongratz writes "Hailing from the world of physics, chemistry, and assorted geewhiz, Lynx Photonic Networks announced a photonic switch with less than 5ns packet switching. "...multiterabit switching systems..." That's what I call bitchin' switchin'." And unlike certain optical switches discussed here before based on bubbles moving in liquid, this variety "does not have any moving parts, nor does it require a change in the physical state of the light signal." 5 nanoseconds.
A few things to think about:
- you need routers at the edge of any optical network to talk to the copper world, and currently the bottleneck is building terabit/petabit routers that will go fast enough (as well as doing added value stuff on the edge such as QoS and VPN processing). Cisco, Juniper, Avici and co should carry on doing well.
- over time, optical switches are likely to become MPLS-enabled, meaning that they have an IP+MPLS (Multiprotocol Label Switching) control plane, and *appear* to be routers even though MPLS is laying down all-optical paths through all these switches. There is a lot of work going on in this area, but ultimately it means that ATM switches, Frame Relay switches, SONET cross-connect switches, and DWDM lambda switches will all look like IP routers. This is a Good Thing, since otherwise the edge IP routers would have to peer directly with all the other edge routers, generating too much routing traffic and consuming too much CPU in each router to be viable (the 'large, flat network' problem).
- Optical Burst Switching is a way of building all-optical switches (at least in the data path) that can at least switch packet trains (bursts). What happens is that an edge device signals ahead of the actual packet burst, on a separate control channel - the delay between the control packet and the data burst gives the electronic control-path part of the optical switch enough time to do the switching (5 ns is a long time at optical rates). This is an alternative to using MPLS or similar to lay down optical paths that are more or less fixed - of course, reliable and fast setup of switching is important to this working well.
Don't forget that these are just the transition times for TTL. You also have to hold the signal high or low for a specified time in order to trigger the next logic gate this one is connected to.
You mention the 74LS00 quad-NAND as an example. Don't forget that this chip exists in a number of different logic families. The 74 prefactor means TTL, and the LS means Low-Power Schottkey. There's also 74F and 74HC and others I can't remember right now that might be faster and/or lower power (albeit with somewhat-varying voltage levels).
We don't use TTL in today's computers of course, it's too slow
TTL is a 'slower' logic family because it runs the BJT's (Bipolar Junction Transistors) in saturation. This is good for power requirements (although not as good as CMOS), but it when a transistor is saturated, it takes some time to come out of saturation for the next cycle. THat's the inherent limitation of TTL.
and chips requiring 5V signals produce too much heat for small circuit paths in the chip.
It isn't just the voltage that causes the waste heat, it's the current too. Remember that TTL signals have low current at 5V, and high current at low signal. Look at the datasheet yourself. (This is for 74ALS00, and a 8-pint SOIC surface-mount device, but functionally equivalent to the DIP 74LS00 you most-likely used in your class).
Typical values in the low state are 0.1mA at 0.35 V, for a power of 35 uW. The high state is 3V at 20uA, for a power of 60 uW. So you see it's not just the voltage that causes the heat.
Because the low states use more power, control signals for gates (for instance, hi-Z output control in tri-state chips) use inverse-logic to activate them. Ie, a signal that will be used only occasionaly will typically have a logic-low activate it's function. Saves power in the long run, and seems kind of weird when you design TTL circuits at first.
Now if you want high speed, look at ECL (Emitter-Coupled Logic). This logic family is really fast. unlike TTL, the BJT's are not run in saturation, so they can switch faster. A side effect of this is that the transistors use more power, and hence run hotter. Like most things, it's a tradeoff. The fastest commercially-available logic family I've seen (and used) is ECLinPS (pronounced Eclipse), for ECL-in-PicoSeconds. These chips can run at several GHz! Pretty sweet. Look here for a datasheet for the ECLinPS NAND gate.
Unfortunately, one of the fastest logic companies has gone out of business about 10 years ago, and I've only been able to glimpse some of their datasheets. It was GigaBit Logic, who in the late 80's and early 90's had logic devices that beat the pants off of what we have now, implemented in GaAs (Gallium Arsenide). However, it cost way too much to develop profitably, and sadly the company is gone. Datasheets had devices listed at 10GHz (although I haven't tested any so I can't guarantee how accurate the datasheets are). .V / _` (_-<_-<
.\_/\_/\__,_/__/__/
__ __ ____ _ ______
\ V
make world, not war
Sounds somewhat missleading to me. While clearly this technology is facinating and will outperform mirrors and bubbles, I raise some doubts about these claims. First, the light signal must be translated into electronic signal in order for the processor to make the switch (because they don't have an all optical processor). Second, they do have moving part in the optical gateway which is heated in order to polarize the light for a particular channel. What is the durability of this gateway? How will it stand up over time? How far can this trick be expanded? Sounds like 64x64 will be pushing the laws of physics.
Someone you trust is one of us.
Put "5 nanoseconds" into perspective.
The speed of light 'c' is 299792458 m/s in a vacuum. A nanosecond is 10^-9 s, so that makes an easy 0.299792458 m/ns. While physics is usually not a good time to switch away from the metric system, that's about one foot per nanosecond (11.8028526772 in/ns).
One foot is roughly one light-nanosecond in a vacuum.
If you have played around with Transistor Transistor Logic (TTL) 5Vcc discrete logic chips in school, you probably learned that gates don't settle their state changes instantly. They take some time to drift from ~5V to ~0V output, or vice versa. The NAND gate is the root of all simple TTL gates, you can implement any simple gate with just NANDs.
One manufacturer's 74LS00 quad NAND package has a "time to pull low" worst-case of 15 ns, and a "time to pull high" worst-case of 22 ns. This is per input bit to be processed. Before that worst-case time has elapsed, you can't be sure you're getting the right answer out of the chip. Managing propagation delays is the biggest reason for providing a CLOCK which drives all the logic at nice clear intervals. The interval has to assume the worst case of any of the involved logic. No wonder the Apple ][ clock was rated at just over 1 MHz, giving ~1000ns between clock ticks.
We don't use TTL in today's computers of course, it's too slow and chips requiring 5V signals produce too much heat for small circuit paths in the chip.
This article is saying that a packet of information sent into the switch as a beam of light can be switched intelligently from its current course to some other course in the time it would take that packet to move just five feet(*).
(*) 'c' is the speed of light through a vacuum; a fiber isn't vacuum; I know that.
[
Sorry. Had to say it.
But seriously... it's old stuff, but many may not be familiar with George Gilder's interesting articles, particularly Into the Fibersphere, on the implications of really, really fast networks. Like the notion that computers may become the bottleneck in the network, and that a packet would be better routed to the other side of the world and back through pure fiber, rather than though the computer next door and back. And how compression becomes passe when it's slower to decompress something than to send the thing uncompressed.
Interesting observation: when computing power was expensive, programmers were paid to conserve it, writing very tight assembly code. Now that it's cheap, programmers are expected to "throw switches at the problem". But bandwidth is expensive, so they write to conserve it. On an all-fiber network, they may be expected to "throw bandwidth at the problem".
Lots of good stuff here, especially considering it was written in 1995. Hope you like it.
"The best we can hope for concerning the people at large is that they be properly armed." - Alexander Hamilton
Maybe if they used these fully optical transistors it might be faster. However, I have slight doubts about the speed of these babies because of a chemical reaction that is central to their operation. But then again, ordinary silicon transistors are based on the diffusion of electrons, which is slow as hell compared to something purely optical, so these might well turn out a lot faster.
--
Escher was the first MC and Giger invented the HR department.
The only thing I hope about all this new technology is that advantages should always include: 1. Lower power consumptions, 2. Smaller footprints, 3. More recyclable material. I'm not really an environmentalist but I think at this point in time we should all become a little concerned with all these new gadgets that are taking increased amounts of power.
yoink
BTW, 256 bps = 16 hex digits a second.
Not last time I checked... 0xF = 1111b = 4 bits; 256 / 4 = 64 hex digits per second.
--
BACKNEXTFINISHCANCEL
Here is a link to their related patent with some more info on the tech used. Pretty damn cool.
--
Jon - TheSpork
"...and I've developed a program that downloads porn from the internet a million times faster!" -nerd from simpsons
...And switch packets in five seconds
...And compile mass libraries of MP3s and warez despite industry controls
...And put giant arms on space stations that make the space shuttle arm look like a grasshopper leg
...And sequence the entire human genome
...And make god games that can seriously damage your view of reality.
Why the HELL am I still getting emails that start like "Looking for HOT, HORNY Teens? Look no further!!!!!!"
The next Slashdot story will be ready soon, but subscribers can beat the rush and slashdot the links early!
Concerning power consumption, it is easy to figure out for yourself that this will dramatically lower the required wattage, since the period over which power is needed will dramatically be lower for each switched packet.
Less obvious, but potentially smaller footprints are needed in such a system, since smaller switch times require smaller curcuits (or whatever...)
Guess the recyclable material will not be an issue here, since this type of equipment will likely be high-end industrial, not for your $20,- ethercard.
Lynx's switch is really a circuit switch and not a packet switch. Let me explain why.
The term Packet Switching commonly refers to statistical multiplexing of packets onto a lower layer channel like say a single WDM channel or a SONET stream. For statistical multiplexing, you need buffers at input or output ports because if two packets arrive at the same port at the same time, one of them has to be buffered while the other is being transmitted. Thus, the common usage of the term optical Packet Switching implies optical buffering and optical processing available in the switch. In other words, its like an optical implementation of a packet switch like an IP router or a cell switch like an ATM switch.
In contrast, these "photonic switches" in the market like those of Lynx are like circuit switches (like TDM swictches used to set up circuits in telephony.) There is no buffering and statistical multiplexing and intelligent forwarding features. The switch may of course still have to do opto-electronic conversion and look at the "packet" or "frame" headers to determine the incoming and then outgoing port numbers. But remember that this makes your switch dependent on the bit rate, clock timing and protocol format.
In contrast, pure wavelength switching (WDM switching) as opposed to this all-optical "packet switching" is totally independent of bit rate, clock timing and protocol format. You are switching light wavelengths and not packets here. This is one of the major advantages of WDM networks since their capacity can be dynamically upgraded in response to customer demand by just upgrading, i.e. adding more wavelengths or increasing bandwidth per wavelength channel - you don't have to go visit every node in your network core and upgrade all the equipment. This advantage is not available if you made your network out of these optical packet-switches like those of Lynx.
All said, wavelength switching and optical "packet-switching" are not necessarily competing technologies. The former is more suited to backbones and long haul networks..while the latter is more suited to local and metropolitan areas.
Note that I wrote "packet switching" in quotes here.