Slashdot Mirror
Optical Memory Could Speed Up the Internet
ananyo writes "Bits of data travelling the internet have a tough commute — they bounce back and forth between optical signal lines for efficient transmission and electrical signal lines for processing. All-optical routers would be more energy efficient, but their development has been hindered by a lack of optical memory devices. Now, researchers have developed just such a device (journal article abstract), paving the way towards a faster, more energy-efficient internet. The devices are based on optical cavities that can be switched between light-transmitting and light-blocking states to construct digital signals. Researchers have been working on such devices for several years, but previous versions used too much power and could not retain data long enough. The new memory cells use just 30 nanowatts of power, 300 times less than previous designs, and can retain data for one microsecond — long enough to support processing."
(See also this paper on all-optical swtiches by four of the same authors.)
As most of the time in long-haul connections is actually spent traveling, this sounds like nonsense to me.
If you have a faster processing of a packet then the router can begin work on a new packet, this may help in the speed looking at the whole picture not just the single connection.
actually no, the travel time is much shorter than that of processing. in fiber-optic networks, latency appears on the nodes.
my sig pwns your sig
So the key is that the medium is able to change its refractive index sufficiently so that there is total external reflection apparently (0) and (almost) complete transmittance (1). Thus, the medium's optical properties (index of refraction which is ultimately a measure of the speed of light in that medium due to the material's permittivity and permeability) dictates its nonvolatile memory applications. You change the medium's optical properties itself with a "write" laser.
The "read" laser (which they call bias but is a bit confusingly used to me) allows you to read off the "memory value" (really just transmittance as a function of the index of refraction set again by the "write" laser).
So the power consumption comes from using two lasers. So it makes me wonder, can you cut down the power requirements by using an LED with a monochromatic wavelength filter? Sure it won't be very efficient in getting a single wavelength, but perhaps you don't need that much optical energy?
Plus there is nothing in the optical realm that even approaches the processing power of an electrical based CPU. Optical routing, even with this very transient optical storage ability is still quite a ways off.
The researchers seem to have missed the huge leftover stock of photographic film. Stopping it completely provides indefinite storage while spinning it 1 mega frame / second satisfies microsecond processing needs.
Actual link latency doesn't effect throughput, while all the time spent going from optics->CPU->optics does. If you could build an all optical router, it would be theoretically capable of very good results. Also, fiber is frequently used for much shorter hops as well. We have hundreds of optical links at the place where I work, and that's all inside a very small room. If more of the infrastructure was 100% optical, that could improve things a lot over the next few years.
Umm, do you know how fast the speed of light is? It's not speed that is the issue, it's the time requires to process the light which is the issue. If it takes 5ms to process light, that means that light has to be made in pulses of 5ms or else signal is lost. Longer pulses = longer travel time in a way. That is why processing purely in optical extremely important as conversion between optical to electron and back is slow in comparison to a pure optical router. When you have lots of routers between endpoints, speed is basically reduced down to switching speed which is the true bottleneck. Remove the switching and you remove the bottleneck. Now, it won't remove the switching speed at the endpoints but it does reduce the latency from switching that occurs in between.
The article skips over the issue of how many bits they can store. It does indicate that numbers > 1 have been achieved, but RAM in megabit, let alone gigabit size, seems to be a long way off.
There have been a few optical switches with fiber optic delay loops. If a packet comes in and the outgoing link is busy, the packet is shunted to a delay loop for one packet time. This works best if the packets are all the same size, like ATM, but it's been made to work with variable sized packets. So far, there's not much commercial technology in the area. Lots of papers, though. People have been working on this problem for over a decade, and there's a little progress each year.
A few bits of pure optical storage and logic will help. If there's enough to handle packet routing and tags, a useful switch can be all-optical, even if storing the data packets themselves in "optical RAM" isn't feasible.
speed of light in fibre is about 0.6-0.7c. If we assume 0.7c, in 10GE, 1500B IP packet is about 260m long.
Normal router (not low latency switch, which is order of magnitude faster) takes about 10us to process packet, this is equivalent of 2100m of fibre.
So if you consider transatlantic link or even use intracity links of 100km or so, time router spends processing is almost negligible to the observed RTT, it is only limited by speed of light.
Use case for optical switching isn't about reducing latency, it'll make some electrical engineering problems easier to solve.
A) electricity does not always propagate at the full speed of light.
B) assuming it does, it is still a significant latency factor-- ping times of 200ms from New York to Shanghai are reasonable (depending on your ISP), and speed-of-light delays would contribute 50+ms to that each way assuming a straight westbound path (which is unlikely). (Note that it IS shorter to fly north instead of west, but that is not how the cables are laid).
If you were to ping all the way into a local chinese endpoint, you would see significant latency occur at the transition from US routers to Chinese routers, but I assume that is because of the crazy IDS / DPI stuff they do.
you're confusing latency for throughput
So...anyone got any super-fast optical memory?
They can go to hell. Where was this done? Did they use ANY public funds? If so, i already own it, so give me the damned article.
---- Booth was a patriot ----
The speed of light is a huge issue.
Try to design a system for a synchronous data replication from, say, New York to San Francisco. You can't without killing performance because the round trip time between the two for fiber or copper connections is 40ms. Even with 0 switching time, the speed of light is too slow to do such a thing.
Fiber optic or not submarine cables still need optical repeaters every 100km or so. Moving these to optical switches would allow for easier deployments.
this dude is FUCKING smart as shit!
I can't see optical memory being a big improvement by itself; you'd still have to switch to electrons to go through the processor, then convert back to photons for transmission.
Now optical processors, that would make a BIG difference.
This is informative and insightful. Google agrees: (2100 m) / (0.7 * c) = 10.0069229 microseconds. Gotta love their calculator, but one wishes it also provided accuracy in the results that got something to do with presumed accuracy of the constants in the expression. As given, the result should have no more than 2 digits of the significand shown. The way things are, we get grad engineering students who mindlessly put those useless digits in the reports, exams, etc. Never mind all the undergrads who seemingly have a digit fetish (at least in the U.S.).
A successful API design takes a mixture of software design and pedagogy.
it does not change the fact that my internet connection is on shared bandwidth copper for dozens if not hundreds of miles, hitting a 20$ netgear router
increase speeds by 10 fold, doesnt matter to me, cause the price to home service is retarded and I will not feel a single bit of it ... or in other words get your shit straight
The speed of light in an optical fiber is about 60-70% of c. This can be divined from fiber's refractive index (1.50). Or you could just Google it.
So light travels at about 200,000,000 m/s in an optical fiber. That would make the propagation time to encircle the equator (40,000,000 m) about 200 ms. NYC to Beijing (11,000 km most directly, 14,000 km westerly) around 50-75 ms each way. Overall, propagation time in long-haul situations is still about 1/2 the total latency. But there are lots of situations that use fiber in the short- and medium-haul, and oftentimes several hops along the way, where the propagation time is negligible compared to the time lost in switching.
Right. 0.07+0.07 ~= 0.1 and 0.11-0.022 ~= 0.088.
That's what we've been told so it must be correct.
Switching time is generally on the order of microseconds. Routing time COULD be longer, that depends. Number of hops to local destinations is generally going to vary between 4 and 8 hops, depending on ISP.
For local destinations, latency should be around 15-30ms. The majority seems to happen when you shift from one ISP to another; I dont know the full reasons. If you are staying all within comcast, your latency can be astonishingly low. (10ms)