MIT Uses Machine Learning Algorithm To Make TCP Twice As Fast

An anonymous reader writes "MIT is claiming they can make the Internet faster if we let computers redesign TCP/IP instead of coding it by hand. They used machine learning to design a version of TCP that's twice the speed and causes half the delay, even with modern bufferbloated networks. They also claim it's more 'fair.' The researchers have put up a lengthy FAQ and source code where they admit they don't know why the system works, only that it goes faster than normal TCP."

Comments

I'd be wary.

[Fishchip]

This is how things like Skynet get started.

Re:I'd be wary.

[Sqr(twg)]

That's easy to prevent.

Re:I'd be wary.

[jamesh]

At the very least i'd be doing a grep for things like "kill all humans" in the source code.

Re:I'd be wary.

[jma05]

That's mostly unnecessary

Re:I'd be wary.

[Zalbik]

. If we do start living in a work where AI's are creating things that we don't understand, but use anyway, then everyone should be concerned.

That certainly would be scary. But that's not the case here. Researchers certainly understand the algorithm; they just don't understand why it improves TCP performance so much. There is a huge difference.

A lot of math is like that (we understand something to be true, but can't prove why). However, I'm not concerned that those algorithms are going to jump of the page and throttle all of humanity in it's sleep.

Uh Oh...

[Anonymous Coward]

they admit they don't know why the system works, only that it goes faster than normal TCP

And so it begins...

Re:Uh Oh...

[Richy_T]

If they don't know how it doesn't work then they don't know when it won't work because of some unaccounted for situation. For an example, consider that pedestrian bridge in London. Engineering is about more than designing for the common case. You also have to consider outliers.

All Jokes Aside... Still No.

[Jane Q. Public]

Allow a computer to design a faster TCP? Sure!

Let them actually implement it without knowing how it works? Oh, Hell no!

I'm not talking "Skynet" or anything here... but if it breaks, who's going to fix it?

Re:All Jokes Aside... Still No.

[Anonymous Coward]

FYI: There's a difference in "knowing the precise mechanism for how it works" and "knowing that the algorithm is stable" are two very different things. Presumably they've proven the latter.

Re:All Jokes Aside... Still No.

[RedBear]

Allow a computer to design a faster TCP? Sure!

Let them actually implement it without knowing how it works? Oh, Hell no!

I'm not talking "Skynet" or anything here... but if it breaks, who's going to fix it?

If it breaks can't we just fall back to the current inefficient algorithms? With the performance and fluidity improvements promised by this approach it could be hugely beneficial to all kinds of networks, even if no one yet fully understands why it works better. They'll figure it out eventually.

Re:All Jokes Aside... Still No.

[Intropy]

We're already in that boat. One of the reasons it's so hard to make changes is that nobody really knows why the internet works. We know how and why individual networks work. We can understand and model RIP and OSPF just fine. And we know how BGP operates too. But that large scale structure is a mess. It's unstable. The techniques we use could easily create disconnected or even weakly connected networks. But they don't except for occasionally a single autonomous system falling off. We've built ourselves a nice big gordian knot already. We know what it's made of, and we know how it operates, but good luck actually describing the thing.

Re:All Jokes Aside... Still No.

[afidel]

Meh, it's like the AI designed antenna, we don't have to know WHY it works better, just that it does and how to build a working copy.

Re:All Jokes Aside... Still No.

[Clarious]

Think of it as solving a multiobjective optimization problem using heuristic algorithm/machine learning. You can't solve the congestion problem completely as it is computionally infeasible, now they just use machine learning to find the (supposedly) optimal solution. Read TFA, it is quite interesting, I wonder if we can apply that to Linux writeback algo to avoid the current latency problem (trying copying 8 Gb of data into a slow storage medium such as SD card or USB flash, prepare for 15+ seconds stalls!), the underlying is the same anyway.

Re:All Jokes Aside... Still No.

[Clarious]

A bit offtopic, roughtly 10 years ago I came to /. and was amazed by the technological insight/information in the comments here. And now more than half of the comments are jokes about skynet without any insight of understanding what TFA is about. Of course, informative posts still can be found often, but slashdot has fallen quite low...

Re:All Jokes Aside... Still No.

[techhead79]

we don't have to know WHY it works better, just that it does and how to build a working copy

But the fact that it does work better means we're either missing a part of the picture that is obviously important or the AI version is leveraging quirks with the system that no current model we have represents. I'm shocked to read that anyone would be comfortable just ignoring the why of something just so we can progress beyond our understanding. If we don't understand the why then we're missing something very important that could lead to breakthroughs in many other areas. Do not let go of the curiosity that got us here to begin with.

Re:All Jokes Aside... Still No.

[blankinthefill]

I don't think they just drop the questions and run with it. I'm pretty sure that, when we don't understand how things that are useful work, we just implement them... and study them at the same time. I guarantee you that SOMEONE, at least, is studying why an AI antenna works better than our man-designed ones, and they're doing it for the very reasons that you mention. But I think the point the GP was trying to get at is that we've never let out ability to not understand things hinder our adoption of those very things in the past, and as long as we have good evidence that this thing performs correctly, and we can replicate it, then why wouldn't we use it at the same time we study it?

Re:All Jokes Aside... Still No.

[Ichijo]

So we built a computer that figured out the answer. Now we just need to build an even bigger computer to figure out the question!

Re:All Jokes Aside... Still No.

[Animats]

One of the reasons it's so hard to make changes is that nobody really knows why the internet works.

We still don't know how to deal with congestion in the middle of a pure datagram network. The Internet works because last-mile congestion is worse than backbone congestion. If you have a backbone switch today with more traffic coming in than the output links can handle, the switch is unlikely to do anything intelligent about which packets to drop. Fortunately, fiber optic links are cheap enough that the backbone can be over-provisioned.

The problem with this is video over the Internet. Netflix is a third of peak Internet traffic. Netflix plus Youtube is about half of Internet traffic during prime time. This is creating demand for more and more bandwidth to home links. Fortunately the backbone companies are keeping up. Where there's been backbone trouble, it's been more political than technical. It also helps that there are so few big sources. Those sources are handled as special cases. Most of the bandwidth used today is one-to-many. That can be handled. If everybody was making HDTV video calls, we'd have a real problem.

(I was involved with Internet congestion control from 1983-1986, and the big worry was congestion in the middle of the network. The ARPANET backbone links were 56Kb/s. Leased lines typically maxed out at 9600 baud. Backbone congestion was a big deal back then. This is partly why TCP was designed to avoid it at all costs.)

Re:All Jokes Aside... Still No.

[Anonymous Coward]

> I'm shocked to read that anyone would be comfortable just ignoring the why of something just so we can progress beyond our understanding.

ML often works like that.

You put the inputs into a function... it spits out a model. The model can be considered as an optimal orthonormal basis for the space it was encoding, but its REALLY REALLY hard to understand the dimensions that basis is in. Sometimes, you can take an image categorization model and see "ah, this is the blue shirt dimension. It seems that people wearing blue shirts are far along this axis"... but most times, you have NO IDEA what the model is capturing.

Re:All Jokes Aside... Still No.

[jkflying]

It's not that we don't understand *why* something like a genetic-algorithm designed antenna works so well. We can evaluate its performance using Maxwell's equations and say, "Yes, it works well." without ever having to build the thing. What we don't have is a set of guidelines or 'rules of thumb' that can result in an antenna design that works just as well.

The difference is that the computer evaluates a billion antennas for us, doing some sort of high-dimensional genetic optimisation on the various lengths of the antenna components. It doesn't 'understand' why it gets the results it does. We do, because we understand Maxwell's equations and we understand how genetic optimisation works. But Maxwell's equations only work for evaluating a design, not for giving a tweak which will improve it. And we're dealing with too many variables that are unknown to have a closed-form solution.

As for this algorithm, they basically did the same thing. They defined a fitness function and then repeatedly varied the code they were using to find the best sequence of code. However, unlike the antenna analogy, they used actual equipment to evaluate the fitness function, not just a model. This means that they don't have an accurate model, which means that your complaint that we don't know why this works is entirely valid, and the antenna analogy is not =)

Re:All Jokes Aside... Still No.

[Clarious]

It is not that simple, take flash memory for example, if the blocks are erased then the write will be very fast, but the write speed will slow to a crawl if they aren't. You can't predict the writeback latency at all, you can only (heuristically) adapt to it. As for the GNU/Linux's complexity, I don't think there is any problem with it, most IO operations are cached in memory, only when you need to flush it down to storage medium then the latency problem appears. I have read somewhere that Linux is optimized for throughput workload (for big server), so the desktop users have to suffer, for them responsiveness is more important than throughtput.

Re:All Jokes Aside... Still No.

[pedestrian crossing]

The jokers are the ones who didn't read TFA. So sad. Especially when TFA is such a good read and is actually "News For Nerds"!

Re:All Jokes Aside... Still No.

[seandiggity]

We should keep investigating why it works but, to be fair, the history of communications is implementing tech before we understand it (e.g. the first trans-Atlantic cable, implemented before we understood wave-particle duality, and therefore couldn't troubleshoot it well when it broke).

Let's not forget this important quote: "I frame no hypotheses; for whatever is not deduced from the phenomena is to be called a hypothesis; and hypotheses, whether metaphysical or physical, whether of occult qualities or mechanical, have no place in experimental philosophy."

...that's Isaac Newton telling us, "I can explain the effects of gravity but I have no clue WTF it is."

Re:All Jokes Aside... Still No.

[Daniel Dvorkin]

I'm shocked to read that anyone would be comfortable just ignoring the why of something just so we can progress beyond our understanding.

If you insist that we know why something works before we make use of it, you're discarding a large portion of engineering. We're still nowhere near a complete understanding of the laws of physics, and yet we make machines that operate quite nicely according to the laws we do know (or at least, of which we have reasonable approximations). The same goes for the relationship between medicine and basic biology, and probably for lots of other stuff as well.

If we don't understand the why then we're missing something very important that could lead to breakthroughs in many other areas. Do not let go of the curiosity that got us here to begin with.

I don't think anyone's talking about letting go of the curiosity. They're not saying, "It works, let's just accept that and move on," but rather, "It works, and we might as well make use of it while we're trying to understand it." Or, from TFA: "Remy's algorithms have more than 150 rules, and will need to be reverse-engineered to figure out how and why they work. We suspect that there is considerable benefit to being able to combine window-based congestion control with pacing where appropriate, but we'll have to trace through these things as they execute to really understand why they work."

Re:All Jokes Aside... Still No.

[Tagged_84]

Oh thanks! I was wondering what time it was here on Earth!

Re:All Jokes Aside... Still No.

[Rockoon]

You should go talk to Intel or AMD about your opinions on the matter, because I assure you that the specific layout of their chips is based on machine learning algorithms. No human can realistically optimize circuits containing a billion transistors.

As a matter of fact, I recall genetic algorithms being thrown at rather small circuit design problems and producing solutions that were better than any human had come up with. Ah yes, here it is: Sorting networks and the END search algorithm.

-- "Even a 25-year old result for the upper bound on the number of comparators for the 13-input case has been improved by one comparator"

Re:All Jokes Aside... Still No.

[Dr_Barnowl]

They did this experiment with genetic algorithms with FPGAs, and it produced a frequency discriminator with far, far fewer gates than any human design.

There was a region of circuitry in the middle not connected to anything, so they figured it was redundant and removed it. The circuit stopped working. Stuff like this can discover emergent properties of systems you never even designed in.

Re:All Jokes Aside... Still No.

[AK Marc]

It's a mystery because in practice, thousands of sessions being tracked is too hard to deterministically determine in a simple static manner. So we use WRED instead. This is WDED - weighted deterministic early detection. What we don't understand is how this does so much better than random drops, mainly because math is hard. Someone could probably take this, and write a mathematics thesis on this. Determining how to drop packets to keep a minimum queue size and have the lowest impact on performance is something that has been worked on for years. This isn't unknowable, or ever really that hard. It's just different and complicated (within a small area of interest, so less than 1% of the population knows what WRED is, let alone how this is essentially an improvement based on it (at least what I could tell from FTA, as I haven't had time to read the source, let alone understand it.

Re:All Jokes Aside... Still No.

[evilviper]

A bit offtopic, roughtly 10 years ago I came to /. and was amazed by the technological insight/information in the comments here.

Yes, all the comments about pouring hot grits on a naked and petrified Natalie Portman had really superb mathematical proofs backing them up...

FWIW, you're absolutely correct. Before /. tried to become digg, and then reddit, and then a flamefest of AGW/evolution/etc. supporters and deniers, there was a much more vibrant community with a tremendous number of experts in fields from mathematics and physics to biology and psychology, always chiming in on the topic of the day, and providing incredible insight into the field and the specific topic that one wouldn't find anywhere else.

It seems that model didn't result in enough ad impressions and profits for the parent company, so flamefests it is. /. has only recently backed off of editors posting complete and total crap, so my belated plans to drop this site entirely were aborted, and I remain. These days, there really are only a handful of folks who provide real insightful comments across many articles. It's easy to spot them if you read this site regularly, and it's such a small group I could rattle off a list of names from memory.

The only reason /. has any relevancy today, and the audience hasn't completely disappeared, is that all other tech sites have HORRIBLE comment/discussion systems that make it hard to follow the discussion, and do not really promote good comments to a wider readership than the first-post crap.

Re:All Jokes Aside... Still No.

[Immerman]

And he would have been right, except some cost-cutting pencil pusher decided to fill a dirigible designed to use inert helium as the lift gas with much cheaper but highly flammable hydrogen instead.

Basically:
Helium airship - cheaper to build (no fire safety measures), expensive to fill
Hydrogen airship - expensive to build, cheap to fill

So they took the cheap route for both building and filling, what could go wrong?

And even as it was the inclusion of a few hundred pounds of evacuation zip-line would have saved most lives, it didn't exactly come down quickly.

Re:All Jokes Aside... Still No.

[Immerman]

The canonical example is we have no idea why we're capable of logical thought, yet that doesn't in any way impair us form using it.

In fact when it comes to most complex systems (economy, ecology, etc, etc, etc) we don't *really* understand how they work, but we muddle through as best we can. Generally speaking when faced with a complex system we take one of two routes:
* Oversimplify our model of the system so that we think we understand it and work from there (the "professional" approach, which often ends catastrophically when those oversimplifications come home to roost)
* Trial and error (the historically far more successful approach, and the only one used by evolution)

Something like the bent-wire antenna with incredible omnidirectional properties is a great example of this: It's not that there's some magical features we haven't discovered about radio, the thing was designed by genetic algorithm within a computer simulation that was 100% limited by our existing models of antenna behavior. But a 10-wire antenna allows for phenomenally complex interactions between those behaviors, and the trial-and-error approach allowed the algorithms to home in on an extremely effective configuration within a problem space far to complex to reason our way through.

An even better example would be the nozzles used by some companies to create powdered laundry detergent - they spent a bunch of money on engineers to design a nozzle that would spray a liquid detergent solution in a manner that created tiny quick-drying droplets. Despite the simulations all saying it should work great, it failed miserably. Then they just built a bunch of random nozzles, tried them out, and used genetic algorithms to home in on an effective design. The difference from the antenna process being that they actually made physical versions of the nozzles to test, because the best available simulations were clearly incompatible with reality.

Re:All Jokes Aside... Still No.

[deathguppie]

Interesting, that when I read your post I couldn't help but think that very same though about politics, and social engineering. People seem to think that any solution that looks good on the surface should work, and yet hardly ever bother to measure the result of any given solution. On the whole we understand that things are f'd up but understanding why is simply beyond us. Makes me wonder what solutions machine learning algorithms would come up with given more and more criteria, and if those solutions would seem reasonable to humans or not. Just don't allow it access to Petman..

Re:All Jokes Aside... Still No.

[Immerman]

I have heard claims along that line - something like one of the protective layers was effectively thermite? There seem to be as many theories as there are people making them, but it's hard to argue that the hydrogen wasn't at least an added accelerant.

Personally I blame the television crews for the real disaster. Without them it would've just been a newspaper story about a German airship burning up and killing some people. With the dramatic visuals though it was the death-nell of the airship industry, for no good reason. The sinking of the Titanic was at least as big a disaster and had negligible effect on the oceanliner industry.

Re: All Jokes Aside... Still No.

[Qwade79]

Here you go:

http://www.sussex.ac.uk/Users/adrianth/ices96/paper.html

I first read about this in a Pratchett book - one of the Science of Discworld ones but never looked it up further.

err, can you walk me through it?

[tloh]

they admit they don't know why the system works

I'm guessing the next big revolution in AI is the quest to figure out how to get digital problem solvers to teach us meat heads how they actually figured this stuff out.

Re:err, can you walk me through it?

[krray]

When they [the computers] do, I bet they don't... Because, you know -- They're Made of Meat

Re:err, can you walk me through it?

[osu-neko]

I'm guessing the next big revolution in AI is the quest to figure out how to get digital problem solvers to teach us meat heads how they actually figured this stuff out.

The thing, we already know exactly how they figured it out -- we wrote the instructions they followed to do so. We know exactly how they figured it out, we just don't understand the solution.

Re:For a second there

[cold fjord]

Colossus: The Forbin Project

Re:For a second there

[Kaenneth]

The books go further, I recommended them.

Headline epic fails.

[girlintraining]

This isn't a redesign of TCP. The network is still just as stupid as it was before; It's just that the local router has had QoS tweaked to be more intelligent. By a considerable margin too. Reviewing the material, it seems to me like it's utilizing genetic algorithms, etc., to predict what's coming down the pipe next and then pre-allocating buffer space; Rather like a predictive cache. Current QoS methods do not do this kind of predictive analysis -- they simply bulk traffic into queues based on header data, not payload.

It comes as no surprise to me predictive/adaptive caching beats sequential/rule-based caching. They've been doing it with CPUs and compilers since, uhh... the 80386 processor. TCP/IP was designed before there was much thought being put into pipelining, caching, parallelization, etc. Using modern algorithms and our better understanding of information system design that's come from 30 years of study results in a noticable improvement to performance? Shocking...

Re:Headline epic fails.

[Lord_Naikon]

Huh? Did you read the same article as I did? As far as I can tell, the article is about a TCP congestion control algorithm, which runs on both endpoints of the connection, and has nothing to do with QoS on intermediate routers. The algorithm generates a set of rules based on three parameters resulting in a set of actions to take like increasing advertised receive window and tx rate control. The result of which is a vastly improved total network throughput (and lower latency) without changing the network itself.

I fail to see the relevance of predictive/adaptive caching. It isn't even mentioned in the article.

Re:Headline epic fails.

[wonkey_monkey]

which runs on both endpoints of the connection

Remy typically runs for several hours and produces a congestion-control algorithm that can be implanted into the sender of a TCP implementation, and then run in real-time. Right now, we do not need to modify the TCP receiver.

Re:Headline epic fails.

[Immerman]

Actually everyone already implements it - dropped packets are a standard part of TCP/IP, the protocol explicitly deals with how to handle them transparently, along with out-of-order delivery - because packets 1, 2, and 3 may all take different routes, and there's no guarantee 3 won't reach the destination first or that 2 will make it at all.

They only question is how to decide *which* packets get dropped when congestion overruns the available queues. In fact if I recall correctly there was some research a year or two ago suggesting that large buffers actually significantly reduce the throughput of a congested network because the end nodes aren't notified of the congestion in a timely manner.

Re:Headline epic fails.

[Forever Wondering]

Active queue management [dropping packets] is already being done by most intermediate routers already. It's been done somewhere along the line in getting this reply to you.

TCP is a "virtual circuit" [stream] protocol. In every TCP packet, there are upstream/downstream byte offsets (e.g. "I've sent you X bytes", "I've received from you Y bytes"). These are examined by the end-to-end nodes only, not any intermediate node or its associated buffers. At this level, packets [per se], dropped or not, don't matter--only byte offsets do.

If end node A is sending to end node B, and has transmitted X bytes, but has only received an ack from B for X-n, after a time, A will resend X-n to X again. This happens bidirectionally with B to A as well.

With overly large buffers, B will [eventually] get two [or more] copies of these retransmitted bytes. This just exacerbates the problem, particularly if there are a number of hops between A and B because most of the intermediate buffers become filled with needless retransmit copies [from disparate virtual circuits (e.g. end nodes C and D that are communicating) that have nothing to do with one another other than they must route through the same intermediate node on their way to their distinct destinations].

The TCP virtual circuit retransmit algorithm assumes that some packets will be lost [due to transmission CRC errors] and/or dropped [or duplicated]. The TCP end-to-end retransmit algorithm [and backoff] is orthogonal to the AQM strategies the articles are talking about. But, large intermediate buffers actually confuse/defeat the TCP backoff strategy and make the problem worse--not better.

When the AQM algorithms work correctly, most packets aren't dropped, because each transmitting end node [eventually] syncs itself to the maximum rate it can send through the [entire] circuit without having data dropped/replicated/delayed unnecessarily.

People don't care about "dropped packets". They care about throughput and low latency. If a packet were never dropped, you'd have neither.

MIT is not the Borg

[CuteSteveJobs]

Kudos, but can't OP say "MIT Researchers Keith Winstein and Hari Balakrishnan". Despite the best efforts of their AI labs, MIT is not the Borg. When someone who works for MIT buys an orange juice, "MIT" has not bought an orange juice.

And if they have software that can outcode me, COOL! How many professions are this lax with job security? :-)

The Traveling Salesman Problem

[SpectreBlofeld]

For some reason, while reading the FAQ writeup, the traveling salesman problem sprung to mind. I don't understand either enough to understand why.

Can anyone educated enough in both issues explain if they are similar computational problems, and why?

Come on now

[dbIII]

As complex systems goes there are far worse. Go ask an engineer or a scientist.

Re:Come on now

[Daniel Dvorkin]

As complex systems goes there are far worse. Go ask an engineer or a scientist.

I am a scientist--specifically, a bioinformaticist, which means I try to build mathematical and computational models of processes in living organisms, which are kind of the canonical example of complex systems. And I will cheerfully admit that the internet, taken as a whole, is at least as complex as anything I deal with.

Re:Come on now

[justthinkit]

Several hundred million computers doing pretty much the same thing is not nearly as complicated as a million species each doing their own thing. The internet is more like a single giant organism.

OSPF

[globaljustin]

It's basically a more complex version of Open Shortest Path First.

Depending on how you understand the term 'autonomous system' you can have a lot of fun with the idea. It doesn't *explain* everything about how this works, but it puts it into context, in my mind.

FTA: To approximate the solution tractably, Remy cuts back on the state that the algorithm has to keep track of. Instead of the full history of all acknowledgments received and outgoing packets sent, a Remy-designed congestion-control algorithm (RemyCC) tracks state variables...

So basically it has, in the minds of these researchers, a really, really well mapped 'routing table' it can access faster than regular TCP.

It's a network control algorythm. It optimizes network flow based on user-identified parameters which result in measurable outputs that can give the user feedback.

Network control algorythm.

NOT machine learning (YAMH)

[dltaylor]

Yet Another Misleading Headline

The paper states quite clearly that once the simulation has produced an algorithm, it is static in implementation.

The authors give a set of goals and an instance of a static network configuration and run a simulation that produces a send/don't send algorithm FOR THAT NETWORK, in which all senders agree to use the same algorithm.

While this looks like very interesting and useful research, it has nothing to do with systems that learn from and adapt to real world networks of networks.

Re:NOT machine learning (YAMH)

[Nemyst]

Um, what? It most certainly is machine learning. Your "simulation" is an offline machine learning algorithm which, given input parameters, finds the best algorithm in the situation provided. Machine learning isn't strictly online algorithms, and it most certainly isn't "systems that learn from and adapt to real world networks of networks", which I'm having a hard time even parsing.

Re:The blurb is flat out wrong.

[paithuk]

The blurb says it "redesigns TCP/IP", and the article itself specifically says "congestion control". Which is NOT part of TCP/IP design. Congestion control is a routing feature.

Seriously, it's both incredible how wrong you are with that statement and that somebody rated it as informative. I suggest you read up on the subject: http://en.wikipedia.org/wiki/TCP_congestion_avoidance_algorithm http://en.wikipedia.org/wiki/Congestion_window http://tools.ietf.org/html/rfc5681