Slashdot Mirror
Rerouting the Networks
prostoalex writes "Scientific American looks at a new approach to clearing networking jams, in research funded by the US military. Instead of using routers to route the packets from point A and point B, thus making some hop in the sequence critical for delivering the message, researchers are exploring a new approach called 'network coding.' (Here is the illustration cited in the article.)" Quoting: "[Four researchers] then at the University of Hong Kong published groundbreaking work that introduced a new approach to distributing information across shared networks. In... network coding, routers are replaced by coders, which transmit evidence about messages instead of sending the messages themselves. When receivers collect the evidence, they deduce the original information from the assembled clues. Although this method may sound counterintuitive, network coding, which is still under study, has the potential to dramatically speed up and improve the reliability of all manner of communications systems and may well spark the next revolution in the field. Investigators are, of course, also exploring additional avenues for improving efficiency; as far as we know, though, those other approaches generally extend existing methods.'"
All of that new NSA equipment would have to be re-worked.
I doubt that the taxpayers would approve.
You are being MICROattacked, from various angles, in a SOFT manner.
just lube the tubes.
we'll cut costs by outsourcing to china ;)
the information transmission equivalent of holography. On one level, it seems sensible and obvious. On another level, it's rather novel (at least, the application of holographic distribution of data to a network architecture). At least, I've never heard of it, either in real-world technology research or even (to my recollection) SF technology.
Welcome to the Panopticon. Used to be a prison, now it's your home.
So they had originally started with one channel from A to D and that channel was also used by B sending to C.
They said that it was inefficient because A to D would have to wait at times for B to C.
Their "solution" seems to be to add a SECOND CHANNEL that will transmit a bit AS FAST AS THE PRIMARY CHANNEL and so both messages will get to their destination sooner.
Why not just use that secondary channel to transmit the data in the first place?
Did I miss something?
to design communication systems that can route around nuclear attacks.
Tsunami -- You can't bring a good wave down!
Well it is a published paper. It's not like it's a big secret now, or even patentable. Thank god....
Money is the root of all evil?
Hey, as long as I get my torrents faster, I'm all for it.
When our name is on the back of your car, we're behind you all the way!
http://research.microsoft.com/~pablo/avalanche.asp x
In the article, I need to transmit all three messages to all receivers and the transmitter/intermediate node link has been fabricated as an obvious bottleneck. How often, in a real network, are the receivers going to have that much connectivity and bandwidth? Also, where is the knowledge that all receiver links are up? The loss of any receiver link means that it will get NO messages, since no complete message is routed of any of its links. If I add enough parity data to allow reconstruction of partial messages, I've just consumed a significant fraction of the supposedly free bandwidth and imposed a higher processing cost at the receiver, which is likely to be the a lower-performance device.
Great way to create battlefield targets, though. Take out any one intermediate link and lots of the troops at the receiver end are cut off.
Wouldn't this method mean that both D and C could potentially decode each others messages? If everyone is sent the same clues, then you should be able to, right?
Could be really beneficial to multi-cast networks like they were saying.
From that, they can stage any number of man-in-the-middle attacks -- the least potent but most widespread of which is convincing a clueless electorate^W userbase that they are certain sources of acclaimed truth, and manipulating them for their own narrow or evil ends.
Philosophers write about this in inspiring 5-page screeds like "On Truth and Lies in a Non-Moral Sense" by Friedrich Nietzsche. That theory in itself is the foundation of postmodernism and some more dangerous philosophies as well. Could it be that philosophy applies to computer science?
Anti-Globalism
They could just start dropping 2 out of ever 3 packets. That would give lots more capacity!
Note that this is a distribution network for multicasting. Nothing wrong with that, but it doesn't do anything for ordinary point to point Internet communications. The cable TV people might find it useful, though.
For their system to work, it seems that A has to send to B and C
which blocks C from sending anything
but C must be sending to D at the same time so that the messages can be merged at the "coder" point.
And the coder would then use TWO transmissions to send the "hint" and the merged data.
It's still sending two data streams (one for merged stream and one for hints) when they were originally complaining about how it was a single channel.
Firstly, this might work for P2P, DHCP, home based (l)users, but it would never be functional in a real world business network. For one, lets take into consideration security. How would this network carry IPSec tunnel information. Those packet headers need to stay in tact not come from ranDumb address. Not only that, they're introducing n+r number of failures where n = number of nodes and r = number of receivers. Secondly sequencing... Would be a nightmare. How would each node know sequencing. What happens if one fails, the sender would have to resend to ALL routers since there is no mention of a mechanism to detect which sequence went where in this topology. Finally... Anything that has to do with governments and routers leads me to remember AT&T and the NSA's taps... First of all, I don't want/need anyone managing my traffic nor would I want to configure this nightmare. It reeks worse than IS-IS + OSPF + MOSPF + MCAST combined on steroids... (My CCIE R&S/Security lab)
Infiltrated dot Net
... "from the assembled clues"?
Congratulations. You've just hardwired a rumor mill. Everyone knows how fast those things travel.
"For every right, an equal responsibility..."
Its a pitty that 99% of the internet's traffic is unicast not multicast so this won't actually help. Plus this gets really hard computationally quickly.
There is a good explaination in this reference. Although that paper mostly treats the case of quantum networks (which is ubber cool) it gives the background. The standard example (not explained well in TFA) is the http://en.wikipedia.org/wiki/Network_coding>butter fly network and the wikipedia article does a reasonable job explaining it. Basically, even with the three different routes, there is still a bottleneck if you just reroute information. But by coding the information, you get it all through. The example of the butterfly network is very simple and worth learning -- just google it if the wikipedia entry or the arxiv paper doesn't help.
Don't do routers already 'efficiently devide' amongst the intermediate nodes, if necessary, by monitoring used/free bandwidth? Isn't the effect the same?
I might be wrong; I just though this already sort of happens.
If you mod this up, your slashdot background will turn into a beautiful sunset!
In spite of being a Mensa-smart guy, I couldn't comprehend exactly how this proposed process would work. I'm blaming it on a needlessly inept description by the author. The best I could visualize was something resembling a parity-bit reconstruction process.
I wish someone could translate this into a more comprehensible form.
The top router can only handle 6 outbound connections at 1 bit per second (total 6 bits per second).
The second rank of routers can handle TEN outbound connections at 1 bit per second (total 10 bits per second).
And they aren't adding any time delays for the merging of the data streams.
I think I can see what they're trying to do the pictures. And it doesn't look right to me (but who am I anyway). Their choke point (the first router) needs a LOT of information about the second rank of routers.
In essence, they're turning 3 data streams into 4 data streams. The 3 original streams go out on 3 of the six outbound lines to 3 of the super fast routers which will send them to 30 of the end points. Given that there are only 20 end points, 10 of them will receive two of the original messages.
Now, those 3 data streams are also merged and that merged stream is sent out on the THREE LEFTOVER LINES. Which end up going to 30 end points.
The problem in my example is that it leaves 10 of the 20 end points with 2 copies of the merged messages and only 1 of the original streams to attempt to unlock them with.
Why not just save time and effort and send the original 3 messages out with 2 lines each? That's GOT to be easier than merging them and HOPING that each endpoint gets the correct unlocking keys.
Their discussion isn't about
A to B to C
and
C to B to A
It's about A sending to D while at the same time B is sending to C.
You've left off "D".
And you failed to account for how B would know ahead of time that C would be sending a message. Which fails completely when you try to account for "D" in the equation. You need to account for the packets telling B which points wish to transmit.
In your wireless example, it would be easier to just skip B and have A broadcast its message to all and sundry and then C can broadcast. B and D would pickup whatever was meant for them.
1. A broadcasts to everyone, D receives the message.
2. B broadcasts to everyone, C receives the message.
Only two steps required. A further 33% savings and it includes an additional recipient.
Point A: (sends a picture of a ladybug)
Point B: 'Okay... is it an animal, vegetable, or mineral?'
Coder: 'It's an animal'
Point B: 'Is it... red?'
Sounds efficient!
This is also a hot topic in wireless networks. In multi-hop wireless networks it can enable nodes to forward many packets in a single transmission. A neat paper about this is here.
Also the coding people are going crazy about this too. There were a couple papers showing that network coding is a simple extension of linear block codes.
"route the packets from point A and point B, thus making some hop in the sequence critical for delivering the message"
In a well designed network, nothing is dependent on one hop, no matter where it is. We call this a multi homed network and most ISPs utilize more than one path to a certain area of the net.
Another thing to consider is that link state routing protocols are capable of quickly seeing that a route is down and picking another route from the topology table, and inserting that route into the routing table, in most cases this is very very quick. Inside an ISPs network this would be OSPF, or between ISPs or other providers this would be BGP, both of which are capable of managing outages as long as a backup route has been provisioned, and in the case of a large ISP, you can be sure it has.
So to sum things up, a given network node in an ISP typically has more than one connection to areas that cover certain IP ranges, if one of the available paths goes down, the routing protocol in use will quickly see that the link is down and start routing over the 2nd link.
I fail to see why people think something needs to change, and this article sounds like they are trying to fix something that isn't actually broken.
Consider using the described method instead of routers, which use the ISO/OSI stack. The message uses an inference, in some ways like how striping works in a RAID 5 drive failure; the failed drive can be inferred from data on the surviving striped drives. Not a big deal. Chunks/nibbles of the messages have a low occupancy time within a cloud, but many receivers must evaluate many messages-- not all are intended for them and must be discarded or validated (imagine spoofing in such a concept as described in the diagrams!).
But in the message diagram, there's a talker, and a receiver that gets messages through inference. Between the two, the cloud between them becomes unbelievably saturated with diffuse messages-- the same reason that ATM looks good on paper with its 53-byte packets, but in reality can't deal with traffic jams. This is why MPLS and other deterministic routing methods were invented-- to qualify transmission routes, not just shoot them like a shotgun into a cloud, hoping that the intended receiver gets the message and deciphers it through receive-side heuristics.
This has been done before, and it didn't work. I'd love to hear comments on why this should succeed.
---- Teach Peace. It's Cheaper Than War.
Network coding is far from a brand new idea. It is introduced in a paper by Ahlswede http://pdos.csail.mit.edu/decouto/papers/ahlswede0 0.pdf published in 2000 and has ever since been a very popular research topic in the networking world (http://www.ifp.uiuc.edu/~koetter/NWC/index.html). These "clues" are linear equations where actual packets can be retrieved by applying a gaussian reduction on the equations.
Its most obvious applications are with multicasting where utilization of network links can indeed be increased in an informational theoretical perspective. The tradeoffs are increased CPU load on the intermediate and end nodes. The research so far is a bit from getting into the practical stages but it has promise. As for "99% of internet traffic being unicast"; even though that might be true, one needs to think outside the box. If it turns out that multicasting will be much cheaper than now (multicasting is now in most cases basically multiple unicasting), broadcasting TV through packet networks might become much more efficient and this proportion could change.
Finally, don't forget that research is still in early stages. I believe there are some years, even decades (if ever), until we will see any of this in practice. Maybe it turns out to be useless for computer networks, who knows. Even so, the basic principle might still prove useful for other applications such as routing inside solid state chips.
First of all, this particular algorithm is really only suited for multicasting the same data to many nodes at the same time. It does not help at all with point-to-point unicasting, which is what the majority of Internet traffic looks like. Secondly, the only real trick here is that you can XOR the data together to transmit more information. So if you have equal length messages A, B, and C, instead of transmitting all three to any particular node, you only transmit XORs of some subset of them (A^B and B^C, or A^C and B^C, etc.) and from those XOR'd versions you can then extract the original 3 messages. Magic! Transmit the equivalent of 2 messages but get 3 messages worth of data! Not groundbreaking; not revolutionary; it's actually a pretty old trick. I've thought for a long time about what real world situation this might be practical for, and have come up with pretty much nothing. It's cute, though.
[i didn't RTFA].
Packet loss/retransmits doesn't matter much, IMHO. In GP example, host B could use some explicit out-of-band signaling to tell A and C what packets are used as xor context. Like, via IP options / IP ID field or something.
Essentially what they are using are erasure codes such as tornado
and digital fountains. The problem with these codes are that even
though they near Shannon's limit, they are not nice towards other
forms of data transmission such as TCP (aka bandwidth hungry).
By using such codes the round-trip overheads that are evident in
protocols such as TCP are eliminated.
These codes are mainly used for massive multi-cast and stateless
loss transmissions.
But using these codes doesn't mean you get rid of routers, it just
means you get rid of the need to persist connections over multiple
routers.
Essentially most comms will be over something similar to UDP. But one
must remember that there will always be an inherent need for reliable
comms. As most of these schemes require an initial exchange of
information in the case of digital fountains a seed and a scheme for
the distribution must be communicated to the end user.
Such technology would be great for wireless connections.
Arash Partow's Philosophy: Be a person who knows what they don't know, and not a person who doesn't know.
So while the discussion was about their example (did you notice the part where I said that the numbers in the diagram didn't match up with their example) you decide that you should be talking about something completely different.
Yes, I'm glad you managed to work your way through the simple A B C problem.
It's just a shame that that wasn't the topic of discussion. And the topic of discussion is the problem they're having getting A to D and B to C.
And it seems that they are talking about using additional channels, simultaneously. Which is why we were discussing their example and why I kept pointing out that you were leaving "D" off of your example.
Can we stay on the subject of their example now?
Exactly. Their diagram is pretty and has unused resources in the overly simplified "before" case ... but it does NOT scale when your have multiple senders and not every receiver wants every message.
So each point in their diagram will either have to be aware of every other point (you think routing tables are bad right now) or they have to send MASSIVE amounts of extra data.
Just from their diagram, they went from 3 lines to 6 lines at the first router. Then they fed those into 6 routers each with TEN lines.
That is a total of 7 routers and and 66 lines to deliver 3 messages to 20 recipients (60 transmissions). Or, 10% more than would be needed if you just ran 1 each.
Now add 2 more senders and 20 more recipients. But 5 of those only want 1 message, 10 want both the new messages and the other 5 want all 5 of the messages from all five of the senders.
You end up with congestion just from the packets telling the receivers which streams are merged and which streams they need to unlock them.
And the machines that only want 1 message have to decode 5 messages to get it.
If this works, it could have several implications that is not immediately visible:
:-)
When the data locally, between two nodes, look like gibberish, does it make it harder to charge traffic by the content in it? Like how a provider may in the future, with IPV6 charge you more per megabyte for say, downloading streaming video than for websurfing. Unlike IPV6 encryption, even the sender and receiver identities would be obfuscated.
If network coding ends at my ISP, it could still charge me. Websites could also charge the ISP if network coding is used merely to transmit data, not store it as the article also hints at.
But network coding would end at my ISP simply because there's only one cable between it and me. With a wireless connection, network coding may offer a great advantage if the user is allowed to connect to several wlan transmitters simultaneously. Not only would this put a spanner in the works for charging-by-content, it may also make our habit of being subscribed to a single ISP seem archaic!
After all, it's only fair that the end-users should enjoy the same level of redundance the internet was originally built with
The concept is pretty interesting although the article is disapointing (and illustrate the priciple using another slashdot lile lame care analogy).
However, even though we would design a workable technology from this idea, I expect huge resistance from router vendors but also from some Internet designer at IETF.
Embedding such an advanced function within the network would violate the dogma where Network needs to be kept stupid and most of the function are to be supported by terminals.
Of course, Internet is not the whole network and we could imagine that such a technology wouls develop on military and banking networks. This would mean to recognise de facto that Internet is not fit for all usages and will never be the Network of networks but one network among others.
the usual tradeoff of squashing data and causing it to take up less space (and move faster) - but at the expense of compute complexity on the send or receive sides?
I do NOT want end nodes to have to work harder than they do today. and routers already do their thing very well. adding MORE complexity to save line bandwidth seems silly to me unless you are still dial-up bound.
--
"It is now safe to switch off your computer."
"We are all geniuses when we dream"
- E.M. Cioran
Just to be clear, "mandates suppport" means neither "is implemented" nor "works."
Besides, if you're using a tunneling method, you'll by definition be creating a unicast stream to the v6 endpoint, which is probably topologically further away than the actual source of the content. Now, if you were able to buy a native IPv6 connection from some provider which didn't rely on IPv4 somewhere, you might be in luck. Then again, it would probably be easier to get most providers to implement IPv4 multicast than native IPv6 + v6 multicast...
Need Geek Rock? Try The Franchise!
...you surely mean Repiping.
With my ECIP protocol we were doing that for video distribution in 1996 and 1997 with the largest network of adult video.
www.ecip.com
I called is server based routing, were a cluster of servers would keep tabs on each others status and network communication quality.
How much latency and loss between each node.
Then when a message was to be sent, they could try direct or go around the blockage by reflecting packets off a series of servers.
Packets would also be split of across multiple paths.
Specifically we had 3 T1 lines at our source site. The end was a user on a PC, we had 40 servers in 40 Co_Lo's on as varied and different backbones as possible.
ECIP would carry the video from the main site with live entertainment to the 40 servers.
End users would get directed via HTML to the server they had the best connection with an then livecam video www.livecamserver.com
would be sent to them.
This as also used for several live video streaming events with Arthur C. Clarke. From Sri Lanka and ECIP
was able to get video out over a very bad and over congested 64K line that provided all Internet to Sri Lanka in 1997.
http://www.dnull.com/~sokol/clarke.html
I am always doing that which I can not do, in order that I may learn how to do it. - Pablo Picasso
It strikes me that network coding seems similar to the way information is propagated in the brain.
and I'm faster than ever!
There exists no way of exchanging information without making judgments. --Bene Gesserit Axiom