>> Jin Li: Hello. It's a great pleasure for us to have Dr. Sundeep Rangan from Polytechnic Institute of NYU to come here and give a talk on femtocell, interference coordination and next-generation wireless systems. Dr. Rangan got his bachelor degree from University of Waterloo in Canada and a master and PhD from UC Berkeley. Then he basically did the post-doc with University of Michigan, Ann Arbor. And he co-founded the Flarion Technology, which is a spinoff of Bell Labs. And they pioneered the Flash OFDM, which is the first US use of OFDM in cellular communication technology. Flarion Technology get acquired by Qualcomm. And Dr. Rangan then served as director of engineering in Qualcomm, leading OFDM effort. And recently reason he just joined Polytechnical Institute of NYU. And let's hear what Dr. Rangan can share with us on femtocells and other topics. >> Sundeep Rangan: Okay. Thank you, Your Honor a lot, Jin, for the introduction. This is of course my second time here at Microsoft Research. I was here before with Qualcomm. Now I'm here with a new Microsoft template, Power Point template. But the talk is still on wireless although it's actually today I'm going to talk about femtocells. Now, Microsoft of course is not in the business of producing femtocells itself but the motivation behind femtocells namely a massive increase in the amount of wireless data traffic is a really of interest to everybody involved in information and communications. So that's why I thought it would give a talk about that here. So let me -- I'm going to just start off the talk with the motivating trends behind femtocells. Now, femtocells really offer a way to deploy networks much cheaper and possibly address the issue of this massive increase in data. But there's two basic challenges in getting femtocells to work. One is at the networking layer and the other is in interference. I'm only going to just briefly talk about the networking aspects. Just to give you an idea of the femtocell concept itself, but also the way that the network architecture revolves has some impacts on the interference. The bulk of the talk will be on the way that we handle interference coordination. Now, interference of course is really a central problem in any wireless system, any time you connect more than two devices they interfere with one another. So I wanted to start off with a bit of a historical perspective on the way that interference is managed in cellular systems and to a lesser extent I'll talk a little about WiFi. And in particular what we'll see here is that we really need to revisit with femtocells how interference is managed. And in particular I want to revisit the issue of frequency reuse which is really the central principle behind the way that interference is managed in cellular systems. Now, when we -- the sort of researching aspect will be in these two last subjects here. And what I want to take a look at the interference, I want to break it into two problems. One is just very briefly talk about femtocell and femtocell interference and then femto and macro interference. And if I have time what I'm going to do, this is -- this is all still somewhat of a speculative came. It really needs the conclusion here is validation. And validation actually with applications of actual application. That's where I think the interaction are Microsoft will be interesting. And I'll just talk a little briefly about how we might try to go about validating some of these concepts in a potential test bed. Okay. So let me start off with what we say is the wireless data crisis. So as all of you probably have experienced right now is that we're seeing explosive demand in the amount of wireless data use, particularly with the iPhone, but just on virtual every device is being realized right now could potential benefit from Internet connectivity. And the result of that is just a huge potential increase in the amount of demand that wireless data devices are seeing. So this graph here, which you'll see show up in lots and lots of presentations, is and estimate. Okay, you could take these estimates for what they are. From Cisco. Which estimates as much as like a 38 fold increase from where we are now to where we are just by the year 2014. And you could see that the bulk of that come from a variety of sources. And some of this is fairly recent that we've come to learn is that the bulk of it is coming from mobile video. The actual source of it won't be that relevant for the talk here, but this will give some understanding. It's good to have some understanding of where it's coming from. What was anticipated that would be a big hawk both in the wire line and wireless is P2P. That does consume a lot, but it's actually video that's dominating. Now, actually I used to be at the startup. And when there was a startup these kind of hawky six curves were always great. You would go to a venture capitalist and that's a little point here, said okay, if you give me money I'm going to like expand your money by this huge amount here. But this is actually not what operators want to see because this is the amount of demand, it is not the amount of potential revenue. So what operators are facing here is a situation where there could be an enormous increase in potentially in data but it's not like the revenue from that data can increase that much. How much more revenue are operators going to get from data services, maybe 20, 30 extra dollars for an iPhone subscription a month? Maybe that's about 50 percent increase. Maybe that could even go up to say 100 percent increase. That's nothing like what you're getting here, a 38 fold increase. So what's going to have to happen to bottom line conclusion is you're going to have to decrease the cost of delivering data per bit down dramatically. Usually somewhat orders of magnitude. Now, that's why they call this crisis. In fact I've seen this graph with various things. One graph said wireless data apocalypse and even be more melodramatic and arrow saying you are here. You can [inaudible] maybe that's like too much but it's definitely a situation that all operators are actually incredibly worried about. Now, the only saving grace is that we've actually been through this before of trying to massively increase wireless data in the past. And if you take a look at let's say the last 20 years, you know, going back to about the beginnings of when GSM was just being deployed in Europe, actually the amount of wireless data capacity is actually increased 4,000 fold. You know, depending on how you count the numbers. Since that time. Now, how did -- it's useful when you want to come up with that last 38 of thinking about how to -- where did that capacity could potentially come from? So capacity you could think of that very simple formula coming from three different sources. One source is that you could increase the number of bits per second per base station, right? So that's really this spectral efficiency. Now, all sort of -- or most academic researchers work on that number, all right? So it's just increasing, you know, doing all the optimizations that's my multiple antenna technologies, you know, better turbo codes, maybe better interference mitigation. But all these technologies contribute to this. So how much is all this academic and fancy techniques work and all the mathematics? Well that's not an insignificant amount but seven. Seven out of that 4,000, all right. And that seven is extremely hard to get. If I've been in the standard space where, you know, an increase of what, half a dB, which is maybe 10 percent, is actually very significant. Now, can we get another -- can we get the bulk from this increasing disk? Probably not. I mean right now we're already running very close to capacity. I think even another seven would be extremely doubtful, all right? So it has to come from two other things. Now, the two other things are actually relatively dumb. One is you can just buy more spectrum. And that actually counts for about 10 times 10 of that factor 4,000. But that too is probably by, you know, there's only so much we can buy there. Although new possibilities like things with cognitive spectrum might open up more. But I will talk about that a little at the end. But the bulk of the increase has really just come from buying more base stations, which is, you know, what cellular operators call cell splitting because they're basically taking a large cell and they're putting lots of smaller cells to create smaller there. So that is the bulk. So what's the lesson of this? If we want to get to our factor of 38 to meet Cisco's demand then really we have to figure out a way to build base stations much, much cheaper. And that is what's motivating femtocells. Femtocells are essentially small personal base stations and which are -- so an example of this femtocell is shown here on the right. This is one that is already actually in production. There are several vendors. This is one from Huawei. And they're basically things that look about the same form factors you get with a wireless access point, bulk significant on the dimensions of the top of your desktop. Now, femtocells are actually just part of a trend. This desire to get smaller, cheaper base stations has been well known in the wireless industry. So if you look at a traditional macro cellular tower these things would go maybe 30 meters high with a large number of [inaudible] at the top. I've already been in decrease well before the introduction of femtocells. For example this is what's called a Nokia MetroSite base station. This itself is a few several years old. That's about maybe one, one half meters high here. So these were already going to a smaller form factor. Femtocells are reducing that form factor much smaller and offering a much cheaper way of deploying these base stations to try to get to you those order of magnitude numbers that we need. So the femtocell concept is not just different in the forum factor, the other important part about it is that it's deployed by the consumer in the consumer's premises. So the idea of the architecture just in this sort of cartoon here is that you put a femtocell just like you would put a attach a wireless access point into your home and it would connect through your personal cable modem provider or whatever your local ISP is. And ultimately through a gateway into the macro -into the operators backbone. But the important point about the femtocell is that it's using the operator's spectrum. So you're using the operator's spectrum but the device that -- with a small device that you've put in the backhaul yours. So you can see here where the savings are potentially coming from for the operator. First of all, the cost of the femtocell even if they have to subsidize is a much smaller device maybe about the size you know -- it can be in the order of like one access point. It's a few hundred dollars as opposed to a macro cellular giant tower is maybe -- although the costs have come down, they're still probably about 50,000 or something. The second part is that the actual back haul eliminated. You're subsidizing that yours. The third part is that the actual core network is completely removed. All the traffic can potentially go through the -- through your backhaul and then into the public Internet with only a minor amount of measurement coming in through the operator's network. And then the fourth part is just the deployment of it. A large part of the cost of running a big cellular network is the amount of tuning and the installation of the cellular towers, and that is all eliminated because you're actually deploying that yours. Yeah? >>: So you [inaudible] I can't use it, right? >> Sundeep Rangan: Yeah. But then you have to hand over. But if you could make that so you'll hand over into the macro cellular network. >>: So now [inaudible] very similar to having this WiFi box currently at home it talks to my broad band router and sends everything out of ->> Sundeep Rangan: Okay. Yeah. So that's a good question. So the question is what's really motivating the difference here between -- it's potentially -- it's actually debated there's a lot of people that are saying that the femtocells just won't pick up. Because anyway, WiFi's already a perfectly acceptable situation in the home. So I think that the only real -- and you know, I think that it's not clear to me which one will actually win. But I think if one's going to make an argument for femtocells they're really of two forms. One is that if the wireless capacity in home was to get to such a high extent that even WiFi was getting overburdened, which it isn't the case right now. Few people actually experience any practical data problems with WiFi networks. Then femtocells might. Because they're operating on a licensed background with a much more efficient cellular standard they could potentially offer higher data rights if that's a problem. I don't think that argument is really that valuable from a customer's perspective until you know maybe if everyone starts streaming HDTV or something in their home over WiFi, maybe then it will start to push the limits. The more important aspect is this, that when it's actually outside your private home if this is going to become a model that the operator will provide in public spaces like in what may be called picocells or something like this where it's a -you know, at home maybe you go to WiFi but what do you start to do when you start to roam and try to, you know, go into a suburban minimal or outside the house in a office building in these areas? And then once it has to become managed by the operator, then the femtocells become much more attractive. >>: [inaudible]. >> Sundeep Rangan: Okay. So that's a good point. So what are you saving then? Now look at from the operator's perspective. It can either deploy a whole bunch of WiFi access points or it could deploy a whole bump of femtocell access points. Say the cost of the femtocell and the WiFi access points are comparable or maybe even the femtocells are just slightly more expensive. So maybe -- so what you're buying from getting the municipal WiFi is getting maybe a slightly cheaper device. But what the advantage of femtocells are is because it's -designed from the cellular perspective we'll have much easier coordination with one another and so the cost then can actually dramatically drop from the operator's perspective over WiFi. Also, things like mobility and other stuff are much better operating in a cellular technology. So that's really what the most basic using femtocells potentially are over municipal WiFi. Because the dominant cost to the operators is going to be actually on the management of these devices much more than the actual cost. And anyway, the cost of femtocells is probably can be made comparable with what it is for WiFi. There's other benefits also. You know, not that these can't be solved with WiFi. When you're looking at a mobile solution then you really want things like power saving, roaming, paging, stuff like that. And those things are actually right now they run much less efficiently on WiFi. For example, you know, cellular technology is much, much more power efficient, even if -- even in active mode but then particularly for things like sleep. And so that's -- that's all those aspects. Once you consider a system where you want ubiquitous coverage which is what's obviously becoming the tracking for devices like the iPhone and stuff, that's when femtocells or pee co-cells or things that are operated in a potentially public domain become actually an attractive solution. Okay. Yes? >>: So does femtocell cover larger area than [inaudible]. >> Sundeep Rangan: It will probably -- think of it as a -- in principle it could, but they'll probably run about the same power output as a wireless access point. So probably comparable the distance. Probably get a little bit better range because actually the cellular technologies have -- can operate in much lower SINRs than wireless access point. Wireless access points probably -- you know, the lowest coding rate's around like 1, 2 dB, and because they actually don't handle fading that well they tend to clap out actually at much higher SNRs than that. And cellular technology been designed from the beginning to work into very, very low SNR regime, so you can get a -- push out the range a little bit better. Okay. Also, there's other issues you can -- you know, in a wireless access point the minimum amount of bandwidth you have to transmit is over 20 megahertz and in a cellular technology, you know, because it's been designed to handle things like voice you can access a very small portion of that bandwidth if you're needing a low rate connection. So the range would definitely be better. So you know you could argue that because it's a cellular technology it's more advanced, that you'll get just a general link layer improvement. But again, I'm not so sure that's a big enough selling feature itself to -- for people to adopt wireless to flip out WiFi access point. It's not like people complain that much about the range of their WiFi devices. >>: [inaudible] try to more money out of me because I'm using a spectrum. But if I am building the [inaudible] is not there, I am not willing to pay more because WiFi solves my purpose. >> Sundeep Rangan: Yes, absolutely. And you know I don't really think that -- I mean, operators are trying to do this right now with the way they're deploying femtocells. I don't think it's going to fly, and the numbers are showing it. If you actually tried even a nominal amount for people to buy a femtocell, people are probably not going to buy it. The only case where people are going to buy it is they want to get cell phone coverage for their voice service in home. For some reason their house isn't getting a good coverage so then they put one of these femtocells in. And, you know, that has a limited amount of value, and it's not really the data then is an issue. So I don't think -- you know, I think the reality is that if operators -- this following's to be deployed, that I will have to give it away pretty much for free. But, you know, if you believe the story that it can offer this tremendous amount of increase in data and offload it, you know, capacity operator's networks then they should give it away for free, which I -- you know, if femtocells take off, that's what will probably happen. Okay. So a little more details then about femtocells. So we've talked a little about this. So if you compare sort of macrocells with femto or picocells you can compare them in a number of different categories. So one aspect is just the cell size. So cell size is, you know, in its typical suburban environment is around right now for macro is about half a kilometer, two kilometer site difference although the specs usually allow very large cells for, you know, rural areas, like up to 60 kilometers. Femtocells you're looking at their coverage that's much more like a WiFi area, even if they were to be deployed publicly. So typically less than about 300 meters probably often need much less than that. The deployment -- so the first part where we see the big cost savings is the deployment. So the macro cellular model for deployment is that it's of course planned and managed by the operator. Now, if you've actually been to an operator you know the incredible amount of effort they -- painstaking effort they go to deploying a macrocellular tower. When I say plan, it means the following, that you have a bunch of RF engineers who do a very detailed similar RF simulation to pick where the cell side is going to be placed. Then they have a real estate team which goes out and actually negotiates a contract to actually deploy the antenna somewhere. And then they install their services on that and go through an elaborate process of downtilting the antennas, driving around and testing it to get everything perfect. I've been doing that myself and it's just incredibly painstaking process. The femtocell comes much cheaper. It's deployed much just like you would deploy an access point. So you open up a box, your customer installs it, plugs it in and deploys it wherever the customer wants. So that whole cost of deployment is just eliminated from the operator, which is in a huge part. Even if the operator were to maintain this in principle it will have a much more plug and play kind of deployment model at a greatly reduced cost to the operator. The second part -- the part that is the same though is the on the air itself, that the spectrum and actually the air interface technology for the most part are completely the same as the cellular part. So it's deployed by the customer but it's using the operator's spectrum. Another difference of course is that the macrocell tends to yield multiple sectors. So what it will have, you saw that picture actually of the tower. So what happens is that the towers actually radiate -- they have several antenna elements, usually three radiating like 120 degree arc into each sector. So that's to get improved efficiency. Here to just make it cost is just a single sector on the directional antenna. The backhaul is the second part picture sector. The big cost to the operators is the backhaul. And it was already a cost when they were running just these 3G systems which run maybe a few years ago anywhere running up to only about two megabits per second. The 4G systems and even the later releases, the 3G systems are pushing over 100 megabits per second in principle. And those backhaul costs would be tremendous. And it's not clear where the money's going to come from. If the -- in the femtocell model, that can in principle be entirely released so that it could just go to say a subscriber's local ISP. The RF, in terms of the actual -- the other part that's reducing cost is just the form factor, as I said, and just the overall cost. So if you look at the raid frequency components in a macrocellular tower you're talking about like a 20 -- for usually about a 20 watts per sector, so with multiple sectors it could go up to a hundred watts. And, you know, if you actually see those things, they're just giant like refrigerator cabinets apart, boxes. And the reason is that to get that and pump that amount of power you need whatever power amplifiers, large power amplifiers, like a PA will be about you know maybe this by that big. And it's actually pretty heavy. You'll want another one on the receive side of the low noise amplifier and they stick in a rack. And then there's all sorts of other costs that come associated with that. You need cooling trays, you need very big cables because you don't want to have any cable loss when you're trying to run this kind of power up to your antenna. On the other hand, a femtocell is usually much smaller, .1 watts to around maybe 1 watt at the absolute most. So it's kind of similar to what you get on a wireless terminal. Which means, you know, if you look at that as a single RF chip it comes very cheap, maybe a few bucks that can get installed. So it's just a radically different part. The other part of the cost savings is that the macrocellular environment is usually servicing hundreds of users usually for voice, and that comes with all sorts of complications, complexities. The femtocell is much smaller like a WiFi access maybe around 4 to 32. And 32 is typically -- that number is fairly high. On the other hand you're not sacrificing on the data rate. You still want to get those peak data rates up there. Yeah? >>: So from a user's point of view, would it save battery life ->> Sundeep Rangan: Over WiFi, definitely. On the terminal side it wouldn't because -- well, at least right now, know, I dope -- there are probably fixes in the .11 spec but the .11 spec is just -- it's actually very, very battery intensive for two reasons. Wire actually transmitting it's just a very inefficient actually mechanism to transmit. You're transmitting over very wide band where you have to constantly be doing these chip level searches which actually guzzle batteries and battery powered cell. But then also when you go into a sleep mode it becomes even more expensive in WiFi because cellular technologies have very, very highly optimized sleep for battery life so that you can keep your phone on and it can pick up a call any time. So they have very efficient ways to just wake up for a very small amount of time, check if it's being paged and then go back to sleep. Those mechanisms, they can be added into WiFi, but they're not -- you know, they're not -- they still don't get the same levels of efficiency that you would get. >>: That's interesting but my question actually was [inaudible] WiFi and femtocells but [inaudible]. >> Sundeep Rangan: Oh, sorry. Sorry, sorry. Yeah, you would save batteries to some degree but not much. You would save the battery power in that you would be closer to the base station. So that way would get -- but maybe not that -- it would be significant -- the bulk of -- on a modern smart phone the bulk of the battery drain's actually on the displays more than the -- the bulk of the battery drains off the display more than the RF component. But even on the modem actually at the talk time it would decrease it a little bit by dropping it. But even that -- maybe -- maybe if you're -- I'm not sure what the number is. I'm going to hazard a guess like 20, 30 percent or something like that if you're right beside the access point versus away. >>: [inaudible] coverage to the pee co-cell versus the macrocell? I mean, if let's say picocell really take off [inaudible] coverage space? >> Sundeep Rangan: Yes. So coverage is just a matter of function like the number of -- picocells you can deploy. So your total data rate is, you know ->>: Because picocell may not be planned that well, right, I mean [inaudible] I mean it may create holes ->> Sundeep Rangan: That's a very good point. So what's going to happen is -so the fear is that because of course you're making a tradeoff you put out unplanned deployment and that's what I'm going to talk about in the rest of this talks. Because you're doing unplanned deployment so now you can get many more base stations. But how much is the capacity per base station go down. Is that your question? So that's what we'll talk about in this talk. And the basic conclusion is that there is going to be a loss but if we do the right things that loss is not so bad. And that, you know, that loss is not so bad and we can handle those sort of smaller corner cases where it does -- fails very badly, and in which case then we'll still get a lot of the benefits from these radically cheaper base stations. >>: [inaudible] and maybe some of the data rates including home technology dependent of operator involvement. >> Sundeep Rangan: Yeah. But because it's -- it has to have operator involvement because of these [inaudible]. Although many of these ideas if they are popular from -- if there's enough benefits on the link layer side then people could try to think about new ways to do unlicensed background. I know. If those ideas on the air interface side are actually good, then you could think of new ways to do it [inaudible] benefits themselves. Okay. So those are the benefits. But here's sort of a high level understanding of where the challenges are to get this. And so a lot of challenges are coming from basically the fact that this is almost like an ad hoc deployment. Now, ad hoc of course computer scientists usually mean sort of distributed routing. That's not the sense in which I mean it here, just the sense of unplanned deployments without the careful planning that an operator usually decides. And there's really sort of two or really three ways in which these kind of manifest themselves. The first is the interference environment itself. So you have, as I said, it's an unplanned deployment. But also there's just the interaction between the macro and femtocell to get the -- no one is going to deploy the femtocells on dedicated spectrum, at least not initially because the demand is not there yet. And anyway, it would be very spectrally inefficient to do that. So they're going to be sharing the same spectrum as a macro and the femto. And when you try to do this as a fundamental difficulty here because there's a large power disparity between the macro and femto, and I'll talk about that. The other difficulty though makes femtocells particularly challenging is what's called restricted association. Our key principle in the way that cellular networks are deployed is that the terminals can connect always to the closest base station. And that's what makes the cellular networks work. That may not be the case in the femtocells because you may not want somebody else to connect to your femtocell. And as a result, you have this restricted association, and that creates many of the really difficult adverse interference environment. A second problem just is really just a need for self configuration, as I said. If we're going to get savings, these things have to be just kind of plug and play. And that has issues both with terms of synchronization at just a physical layer because we need for example to -- need synchronization to smooth the handovers between the macro and femto, but then also some kind of synchronization for interference mitigation techniques. Plus you also need other parts which I won't talk about in this thing, just for actually the way that there are -- devices are managed through the operator and so on. And the third part is really just the networking aspects because no longer -- in the previous part the networks were complete under the operator's control. Now it's a mixture of a network partly going through a localized, partly going through a public Internet and then finally the operator, then how do we try to do that. So I'll only talk a little about this. I just want to talk a little bit just about the standardization efforts. Yes? >>: [inaudible] so when you talked about the [inaudible] of femtocell outside the house, is handover a new challenge because now it will be more frequent as I travel? >> Sundeep Rangan: Absolutely. Absolutely. But on the other hand, where you're going to likely have the density of it will be probably in areas where the velocity will be smaller. But definitely the handover is an issue -- the handover is an issue partly because of the -- partly because of the smaller cells so you'll be hang off more frequently but then also the issue that you're handing over over multiple difficult networking domains, you know, potentially handing off over one percent femtocell connect into one ISP into an femtocell into another ISP and so on. I won't talk too much about that, but that's definitely an issue. Okay. So I wanted to do this slide here just because I'm going to be throwing around some of these acronyms. And by the way, I -- unfortunately this field is just filled with acronyms. So if you ever see an acronym that you don't understand, just raise your hand and I'll tell you. So femtocells are really just part of an evolution which you can -and it's just good to see it in the evolution of cellular systems more generally. So generally femtocells -- cellular systems are characterized into what's called you know 2G, 3G, and 4G. So 2G systems generally refer to the first digital voice systems like back into GSM which were narrow band systems. Now, I put a range of data rates. That's actually -- that range is really the beginning number is sort of the original date rates into the first releases of this pack. And you could see all these systems, you know, it's not like it ended its evolution here. They're continuing to evolve. So you know the GSM/EDGE systems now actually get -- actually that number is even old right now. They have like a two megabit version of these -- or EDGE systems. In 3D systems which are systems that are really the first generation of -- really the first data based systems where actually the first actually offered packet services. And real broad band data rates. So those began in around 2000 and, you know, they were originally around about 384 kilobits but now in the later releases we'll get, you know, very high data rates. Finally four generation systems are what's call long-term evolution and these systems are really offering allegedly higher data rights than that and lower latency. Now, I wanted to put this here because a lot of the -- I'm going to talk mostly about femtocells in the context of LTE. And when LTE was actually in the planning stage back in about 2005, a lot of these -- a lot of the benefits that people thought were moving from -- LTEs and OFDM base system as opposed to a CDAA system. A lot of the benefits that people really thought of that LTE would have really not become really that valuable. One of the benefits here was a very high data rate. Well, the very high data rates, first of all UMTS actually was able to increase its data rates within the system itself. And anyway, those very high peak data rates are not really that possible in a macrocellular system, those are only possible in four antennas. How many devices have four antennas. And you know in your situation that you're very close to the base station which is not going to happen. Now the low latency, sure the low latency is good. But anyway, UMTS dropped down its delays down to 30, 40 mill second or maybe that delay isn't that much. Where the benefit though from LTE now has become increasingly realized is actually much more into femtocell context. So and that wasn't something that was necessarily obvious when people started to design the system back, you know, in 2005. So a lot of these benefits you could see are actually very valuable in the femtocell context. Okay. The data rights are not valuable in the macrocellular because you're not going to get them but factual in a femtocell case it is very possible to -- it is very possible to get these very high data rates in that environment. The other aspect of it which is always on IP connectivity and actually more generally just an IP, all IP back all, that actually will see here really facilitates the easy networking between a local and private ISP and a public ISP. And the other part about LTE which we'll talk about later is on terms of the actual interference coordination. It's really in that domain, that LTE really excels over CDMA systems because of a key concept which we will talk about as subband scheduling. So while LTE really was promoted and initially pushed as a way to try to increase capacity in the macrocell environment, it's gains are not that significant. But actually it's in the femtocell or just more generally in a situation what's called heterogenous networks where LTE will really excel, and that's why I want to talk about it in the context of LTE. Just a little bit in terms of dates here. The actual work on femtocells began before the standardization about the while -- while LTE was going on. It was from pretty early on that people started talking about femtocell aspect back in 2006 and 3G in 2007 and 4G. And so as I said now there's increasing amount -a lot of the effort that worked on standardization for future releases of the technology have really focused on femtocells. There are other groups involved and I just put one of them here which is called the FemtoForum. So you'll see some of the references to their work here. All right. So let me talk just briefly about the network architecture. It's not really my area of expertise, so that's why I'll skip over it a little bit. But also just wanted to only talk about just so you got an idea of the femtocell concept. Okay. So this is one of the alphabet soup kind of slides unfortunately so if you don't remember any of these acronyms, don't worry too much. Actually if you see pictures of network architecture for LTE, they're even much more hideous than this. They have like 10,000 boxes of different sorts. I just picked out a few which -- to simplify it. Okay. So the first acronym we're going to learn is the device itself. That's called the user equipment. I put a cell phone. That could be any PDN or it could actually be a wireless laptop or anything like that. Anything that's getting connectivity. The base stations of a ridiculous acronym called E node B. The E stands for evolved because the previous generation was called node B for a reason that makes absolutely no sense to me. But that's the -- that's the name. So this is the macrocell environment. And the set of them is called the EU TRAN is another ridiculous acronym, all right, so it's called the evolved universal terrestrial radio access network. All right. So that's the base station. Everything behind that is the core network. So in the core network in the user plane which is actually where the data is coming from, the two important boxes are called the serving gateway and a packet data network gateway. The packet data network gateway was the key box and that's the L3 point of attachment. So in terms of IP routing that is the first point of connectivity. If you were to do a trace route after your connection is up, the first box you would see is the IP address of the PDN gateway. Everything from there would be really you consider as L2 routing. No what's specifically happening is when data arrives in from the public Internet, it's classified into what's called an EPS bearer, all right. A bearer is something which is specific to the UE, and the UE might have multiple bearers to affect different quality of services for different IP traffic. And it's mapped to one of these EPS bearers. And then all the routing between the PDN gateway and the user equipment is based on that EPS bearer. So none of these other boxes in between here have any concept of the UEs IP address. So it's all L2 routing. Now, of course it can be IP routing in parts of this like this GDB tunnel between the PDN gateway and the base station. But that's actually transparent to this end-to-end connection. That's just an IP connection between the PDN gateway and the base station. All right. >>: [inaudible] some questions? >> Sundeep Rangan: Yes. >>: So in the case let's say the UE has an IP address. >> Sundeep Rangan: Yes. >>: Is that IP address associated with a gateway? I mean basically who [inaudible] I mean since you mentioned that most of the intermediate basically node only have this information [inaudible]. >> Sundeep Rangan: Right. >>: So who has the information ->> Sundeep Rangan: Only the PDN gateway. >>: Only the gateway? >> Sundeep Rangan: Yes, that's right. So you actually -- when you do -- so you get your L2 connection will come up. Then you'll, you know, a Windows devices you'll get a media up. You'll run DHCP here. And then that DHPC server is running in the PDN gateway. So the PDN gateway will give you the IP address. >>: Okay. [inaudible]. Is that only one or you potentially can have several? >> Sundeep Rangan: You have several. So you have one for every type of quality service. So typically you know your Internet traffic will run on say one EPS bearer but then if you were to offer let's say IMS voice or maybe video conferencing or something like that, if it was not running as a best effort service that the operator was trying to deliver, then they would probably deliver it on another EPS bearer to give it say a higher quality of service. >>: And that EPS bearer would be un[inaudible] by the base station and all the basic ->> Sundeep Rangan: That's correct. >>: [inaudible]. >> Sundeep Rangan: So when that traffic gets routed here, it comes with -- it says this is the -- essentially the Mac address, okay, though it's not -- it's a different Mac address than what you would have with WiFi, and this is the bearer IP for that Mac address. So this is the third bearer, this traffic is on the third bearer of that. >>: How is the quality of service of -- end up in those base stations? I mean, let's say the base station is associated with -- can you use each of them running ->> Sundeep Rangan: Right. Because the traffic comes in here marked with a bearer and then the bearer will have some quality of service profile so then it schedule the one track, one bearer's traffic in with relative priority to another bearer's traffic. So it will maintain a cue for every bearer. So actually there will be a set of cues for every UE and then within the UE one cue for every bearer of that UE and then when it does the scheduling it will pick from -- presumably it will contend to favor higher priority cues over lower priority cues. >>: [inaudible] router, Cisco routers. >> Sundeep Rangan: Right. >>: I mean basically I think basically [inaudible] I mean basically people talk about basically [inaudible] and basically [inaudible] is basically [inaudible]. >> Sundeep Rangan: Yes. >>: Say something like that. >> Sundeep Rangan: That's right. >>: Is those basically information available online. >> Sundeep Rangan: Absolutely. You mean for the air link scheduling? >>: Yes. >> Sundeep Rangan: Yes, absolutely. So you can do -- very much you can do those kind of things, they can do maximum weight matching or any of that. So you have a bunch of cues, you have a priority policy on every cue and then you can do it. The wireless scheduling actually has -- is a more complicated typically than the switch scheduling problem because the scheduling decisions you're making at one base station. And that's what I'll talk about in this talk. We'll of some influences on the scheduling decisions here because they could change the interference conditions. So it becomes a more complicated problem. Also because the channel rates themselves change over time. So there's other things you can get like what's called; multi-user diversity, you can try to schedule not just based on priority but when the actual channel is better versus when the channel is worse and you can play these kind of games. All these kinds of games which can happen in scheduling you know wireless link that won't happen in a wire line link. >>: Well, another question, EPS bearer, this bearer, is there any way to [inaudible] on the UE or on the whole side can influence the [inaudible]. >> Sundeep Rangan: Yeah. So that's the part here. So the operator will -- how the traffic will get categorized. Now, the way that it's been done so far is that you know everything from the public Internet whether you go to YouTube or CNN or something like this, they are all just categorized as generally a best effort bearer. In principle, though, and then the operator services if they were to deploy IP based services, let's say like IMS voice or something, they could classify them on to a separate bearer. Now, there's nothing, though, prohibiting the operator aside from things like network neutrality or something like this. They could say they could make a contract with CNN and say, okay, if my client goes to CNN service, I will then -you know, based on the source IP address of the server or something, I'm going to classify that separately and maybe give that higher priority if CNN gives me some money. So that could happen, you know. Or that, you know, maybe it can also from the port number recognize that this is a VoIP application. And if they want to be nice but they probably don't but if they wanted to be nice to say Skype or something, they could probably pick that up from doing some packet inspection and then give that higher priority. But maybe they don't want -- they probably don't want to be nice. Or worse. Or worse. Yeah. That's right. Now, that's actually a whole opening open part because another aspect of not talking about here is a way that these networks is funding -- is probably through a much more sophisticated -- sophisticated model, pricing model with the content providers. Most people of looked at content -- the pricing with the subscribers, but it's not unusually that the content providers could provide a lot of the funding for these networks. And the operators with this system could leverage their ability to provide different quality and services by the way it's -- you know, marked in the PDN gateway. >>: Basically further question is I know in the LTE there's talking about broadcast now, right? >>: Yes. >> Sundeep Rangan: How is that handled? Is any basically providers deploying broadcasting mode into the picture? >>: I do believe -- I do believe it's definitely handled as a whole part [inaudible] that can be -- that can absolutely be supported. And I believe that the broadcast will also go into the PDN gateway here and then map to a Multicast ID much like -- and then there will be some kind of Multicast bearers here. >>: And then the user basically all the users ->> Sundeep Rangan: Can get that. >>: [inaudible]. >> Sundeep Rangan: That's right. And then the scheduler when it comes to the base station to do something intelligent it can try to send the data to the ones with the -- you don't know the channel quality to all the users and then pick one that will make it work for all of them. >>: So are you aware of any providers using that service for -- I mean ->> Sundeep Rangan: Well, LTE itself hasn't been deployed yet. You know, but presumably once it's deployed it will be, because they're already doing that with the 3G systems right now. So like Verizon does some kind of time slotted basis for some things like V cast and stuff go through some kind of time slot so I presume that it will get -- it will definitely get used. But you know, as we were talking about in the morning, you know, how much will be Multicast versus how much will be On Demand. And if On Demand viewing is the one that, you know, operators are really worried about then ->>: I think you answered my question. It was basically how the PDN gateway will classify [inaudible]. >> Sundeep Rangan: Yeah. So it's based on something like a [inaudible] number and, you know ->>: [inaudible]. >> Sundeep Rangan: You know, that's a good question. I think a PDN gateway could serve in probably the neighborhood I'm thinking of like 10s of -- you know, 10s of base stations. But, you know, what will happen in these network architecture would not -- new network architects, with hundreds of megabits per second going potentially to each base station is, you know, it's not clear. And with the femtocell case then the scaling issue becomes total different now. Potentially there are hundreds of base stations for PDN gateway. So then we need something, you know, very different on the network. Okay. So -- yeah. >>: [inaudible] the address translation that the gateway ->> Sundeep Rangan: Oh, the NAT. So if you get -- say in the NAT would probably -- suppose you instead of having a user equipment here you then break this out to have -- so you probably my understanding I would get -- it will probably much like what would happen right now with a cable modem, you would run right now NAT at the -- I guess you would run it probably at your cable modem. So you will get one public IP address from the -- from the operator and then if you want to break that out you can do network address translation into your private network. Into your private network at the -- on the customer's premises. The same would happen here. You'll get from the PDN gateway a public IP -- a single IP address or maybe they can offer multiple IP addresses if they want to do that. And then if you -- in your private -- is that your question? Yeah. So the NAT runs -- the NAT runs into customers. >>: [inaudible]. >> Sundeep Rangan: They will -- the specs all spec out IP V6. Now, how much the operators networks right now are IP V6 I'm not sure. But all these all these run -- all the protocols are compatible with IP V6. They also -- I'm not going to talk about it here but for -- to do the -- you also want to be able to do handovers IP mobility between from non-3GPB to 3GPB and that will -- they'll generally run either mobile IP and that's also supported either -- they'll either run a dual stack IP -- IP V6 or do a proxy mobile IP depending if it wants to be networked managed, and that will also run in -- that also supports V6. Okay. So that was the -- that was sort of a high level picture of the macro network. Let's take a look at what happens in the femtocell network. So the femtocell network what happens is you don't directly talk again with the serving gateway, instead what you go through is what's called a home E node B gateway. So the home E node B gateway is really the connection between the operator's core network and the public Internet, all right. So you're connected here to your home E node B. It's called the home E node B. Remember, E node B is base station, so home E node B all right is your femtocell. So I've kind of drawn it in a little bit different picture here. So there's a home E node B which will be connected to your local ISP which will then go on to the public Internet and then go to home E node B, right? The home E node B you could think of as kind of like a concentrator. What it will do is just concentrate the traffic, concentrate the traffic from multiple home E node Bs. But essentially that connection it's essentially transparent. So really it looks logically like a connection directly to this serving gateway. Now, I put this up just to show you the picture here. The L3 point of attachment is still the PDN gateway. Now, what this means here is that your home E node B when it comes up, it will get an IP address actually from the local ISP. But that is not again -- you don't see that IP address. So your home E node B gets a local IP address from the local ISP will then get a connection to the serving gate -after he gets its IP connection there goes to the -- well, it has a -- it has a connection here to the serving gateway. But again, if you did a trace route, still your L3 point of attachment is here. Now, this actually -- so if you were to ping, all right, the home E node B's IP address, with the way I've drawn it here, you will go through this kind of ridiculous route. Your L3 point of attachment is the PDN gateway, all right. So you will go through your home E node B who doesn't even know that -- because he doesn't expect your traffic that is coming to him, he will just afford it along, la, de, da, through the ISP, through the public Internet, home E node B, then to the PDN gateway. Through the PDN gateway then -- no, it doesn't go back this way. This is a public Internet address so it goes out through here because he's dumb and doesn't even recognize that it's on his network, right. So that's the way that would work. Now, I'll talk about that in a second to solve that problem. But I just want to point out one big difference, though, between this. So aside from this E node B concentrate here, one of the things here that you see between macro cells is this X2 interface. Now, an LTE, they provide two interface. Part of that is just for handover if you go from one base station to another. But the important other aspect is for X is for intercell interference coordination. So there's actually control messages that are going back between the base stations to coordinate the interference side. You could say oh, I have a UE that I'm trying to serve, could you try to vacate this party spectrum because this part you know I need this is [inaudible] and so on. So all that kind of information is there in the X too. That is no longer possible in the home E node B because I just thought it was not scaleable or logically difficult to try to develop and X2 interface. But you've to find the other E node Bs and so on. Home E node Bs. Now, because of that, that will present some challenges for intercell interference coordination. I'll try to address that in the end of the talk. >>: [inaudible]. >> Sundeep Rangan: That's why. Okay. So just to get around that absurd problem of course the big issue that, you know, part of the selling point is to offload traffic off the -- off the operators network. So that, you know, everything has to go through the operators core network that wouldn't be good. So the way that it's time will likely be deployed is with one of these variants. One is called LIPA and one is called SIPTO. Okay. So let's go again. The traditional IP path would be because this is your home UE if you're wondering what that stand for. All right. Normally, as I said, goes through your local ISP through the public Internet and goes there to the cellular core network and that's your first point of attachment. So that's if you want to get from say from YouTube you go this route here. All right? Which is obviously not good from the customer's perspective, it's not good from the operator's perspective either because the -- you're eating up capacity here and that's particularly bad because these can be very high capacity lengths. So instead what they're working on are two different solutions. One is what's called selected IP traffic offload which is to create a connect -- direct connection. Now, to do that though you need an L3 point of attach in the home E node B itself. So they have this version of what's called a local PDN gateway. And they're still being expected out to try to figure out how to do this, the different variants of this, but there's one solution here. That local PDN gateway could also be used for what's called local IP access. So for example, I put it here a, you know, whatever, flat panel screen here. So you could imagine you were trying to stream wireless data from, it may be some kind of laptop on to that screen. So now that could run over this local IP gateway here. The actual way it's going to be solved, there are a lot of tactical issues to try to deal with this. You know, just basically a case of how you do handoff, how you do security and so on. Those are still being worked out. But they will likely, you know, be part of the femtocell situation. All right? >>: [inaudible]. >> Sundeep Rangan: No. Really there isn't. That's the point. Really in principle you could imagine that everybody that would get -- everything could get offloaded or for control messaging. And like the authentication and you know, operator management. But in principle the overwhelming majority can drop it unless it's the operator's providing the services like IMS voice or something like that. Okay. Another issue I'm not going to talk about it much but it's just to give a picture of it is the security. Now, I put this up just because the traditional -- in the traditional macrocell, you know, deployment it's been assumed that everything -the whole backhaul core network, that's everything from the cell towers and all the network element boxes is trusted and secure. And so actually your data run unencrypted, all right? It's encrypted of course when it goes over the air but it's decrypted and is not encrypted anywhere else in the core network. So if you were to tap into a macrocellular tower right now you would read everything into clear. Now, that's not -- not a problem in an operated -- potentially not a problem in an operated network. And it had a lot of performance benefits by not having to do the encryption over the backhaul. Of course that's not possible if the femtocell situation. So generally what they do is they put what's called a security gateway that would access, you know, between the untrusted public Internet and the trusted IP network. And then you run actual security, the IP set over this which is -- so it's actually encrypted at the security and decrypted here. I know, again, there's just a huge number -- actually the have allocution of security is much like what happened you'll see in the -- when they began to realize with WEP based security they added some un-Godly, like 200 pages of spec into the -- you know, if you read the .11, 2007 specs you'll see like 200 pages of specs on security. It's actually had a similar problem in 3 GPP like 17 specs just on all sorts of varieties of problems with security. So obviously I'm not going to address that here. All right. So let's go on to the main topic here which is just on interference. How much time -- I've actually -- so I'll just talk. >> Jin Li: I mean we can basically [inaudible]. >> Sundeep Rangan: All right. So I'll just skip through a couple of slides. So this was the [inaudible]. So the other key problem in femtocells is the problem of interference. So this picture was really taken from one of the specs on a bunch of different interference scenarios that you're having. Of course interference really the central problem is that in any wireless network it's -- you'll have interference whenever there are more than two connections. And there was an interference we wouldn't have a lot of the problems we have, we would just be dealing with point to point links. But we have interference. But femtocells really create particularly challenging interference problems. Deployment is unplanned, the cell may be restricted and there's these issues of the macro-femto overlay. In addition to that, there's a desire to really make the adaptation very fast. So every intercell interference problem could be solved by appropriate adaptation but then the question is you want to do that quickly and you want to do that quickly because the key to the game of using the air link officially is to be able to dynamically allocate resources back very fast. And that's particularly needed for low latency or bursty traffic. I should just point out that really femtocells the issues faced with femtocells are similar for more generally for what is called heterogenous networks, or this term hetnets, which you'll see a lot in the cellular space. And, you know, they're mixtures of femtos, macros but also other elements like relays and then also mixtures of types and spectrums like between the non cellular part. So this is all sorts of issues that operators are deploying -- facing. Now, [inaudible] really you can think of this two ways to deal with interference, at least in a simple way. One is what's called -- one is the traditional cellular method which is power control. All right. And so power control is with [inaudible] it's basically -- here are two base stations and let's say there are two mobiles and the simplest thing is this. That both base stations will transmit on the link at the same time and frequency -- same time and frequency. So they'll obviously when this -- this UE here will see that entire signal as interference and this UE will see that signal as interference. And then the only thing you can try to do to improve the link is try to do power control. So you can increase the power on your link which will improve your quality at the expense of the others by creating interference there. So the only gain is power control. This, for reasons I'll make clear here has become the dominant method for cellular systems and that's because of the case -- actually this is more optimal in the case when the interference is week. And that's become the dominant method for 3G and will continue to be the dominant method for 4G cellular systems. The other type of method is what you could say is orthogonalization, which is actually a little easier to understand. All you do is at the two transmitters transmit on different resources. All right? So well either in time or frequency. So here that the links -- so this blue link comes on one frequency resource and doesn't have any interference from this link. So that has no interference problem but of course what you're doing is you're sacrificing bandwidth that you only get half as opposed to this that you got the full amount of bandwidth. So a lot of the questions then -- and that's really you could [inaudible] viewed that way is really the way that things at .11 technologies work. We'll see that -- what we're going to see is that femtocells really don't fit neatly into one of these two categories. And what's really needed is some kind of sophisticated hybrid scheme, and that's what I'll talk about here. Now, to get an idea of when to use one, it's interesting to just take a look at a couple of different interference scenarios. So let's take a look at -- let's take a look at a few types. The first one to do is look at the interference in a macrocellular situation. So this is obviously -- the interference is very environmentally independent, so all you could do is run simulations and it's accurate and, you know, as accurate as simulation model is. But let's take a look at a few simulations. So one part here is this model here is what's typically used in studies of macrocellular environment. And you -- this is kind of imagine the ideal placement of base station. So this perfect hexagon and then with a perfect sectorization where the cell sectors are perfectly aligned and so on. So that's you know that's one model you know what you could look at the interference situation you could get from here. So that's kind of a model that you would see in a traditional macro network. So let's deviate from that in two ways. The first way as you can imagine is unplanned deployment. Let's just imagine that I deploy cells randomly and let's look at the interference conditions here. The second part is instead of this macro environment let's look at a femtocell environment. So this is the femtocell environment that's used in a lot of the 3 GPP studies. You have a bunch of apartments which are 10 meter by 10 meter and about half of them, let's say, have a femtocell in them. So that red line is one end of that is a femtocell transmitter and the other is a receiver. So that's a link going on in this connection here. Now, the important part about this is this [inaudible] restricted association. So you always connect to the femtocell in your own apartment because that's your femtocell. Even if that might -- even though there's a wall here, it's often the case that that may not be the closest base station. There might be, you know, might be just on the other side of the wall, right. There may be a closer but you still won't connect to that. So let's see what happens. So let's first look at what happens when you make the undeployment unplanned. So one of the common things that, you know, cellular engineers look at is these interference or what are called geometry CDFs. So the X axis here is the SI in our signal to interference and noise ratio. So it's a measure of the amount of interference that the link is hearing. So this is the way you imagine all the links transmit at the same time. You may not choose to do this, but let's say that you did. Right. If all the links transmit at the same time and then you can look at your signal power over the sum of all the interferences that you got. That's the SINR, all right. And now that's a random adversarial because it depends on randomly where you are. And this is a CDF. So if you look here at the degredation just going from this perfect hexagon to this random environment, you see already about a 2 dB loss. Now, 2 dB loss is not insignificant. That's about what, maybe 70 percent. So you've already potentially -- not 70, about 30 percent. So that's about a 30 percent loss in capacity just from the not having a ideally deployed environment. When you go to a femtocell situation, you get a much more radically different interference situation. And it's sort of there's good news and bad news. Here's the good news. The good news is that there are many links actually which have a very, very good SNR. You know, [inaudible] as much as like 20 dB. When the radio probably can't even benefit for more than 25 dBs. So you're about the limit here. All right. That's because of course these apartments are relatively isolated like maybe this link over here is isolated so it gets a very good SNR, all right? So that's many links. But there are some links which are really very bad, all right, down to what may be -- you could say about eight percent or something are down about minus 10 dB. Now, you might think okay, well what's so bad about that, it's just a small number of links. But remember this. Imagine you were using your laptop. You're not going to remember the 60, 70 percent of times that your laptop was really good. You know, one out of 10 times you use your laptop and it completely fails, you're going to remember that. So that's the actual -- that's the issue here. Even if it's a small apparently that's the part that we really need to address. So the big difference you see here is that there's a much large number of links in a much worse interference condition. And that's coming primarily from this fact of the cell selection, all right, that is restricted that you cannot connect to the strongest cell. So this is where we see the big difference in the interference environment is when we do this kind of graph. Now, maybe just to -- what I'll do is I'm going to skip over all of this. Okay. Let me just take a couple of slides here. So let's take a look at where power -- I said there were two methods to try to deal with interference. One is power control and the other is orthogonalization. So when you look at power control has actually been a very effective method in the cellular environment, all right. This is a graph -- this is an algorithm that I worked on but, you know, there's a gazillion algorithms out there and there's many deployed in commercial systems here. So if you look at the rate CDF actually, so when you run these algorithms and you look at the rate CDF, you can see a couple of different aspects. First of all, that the rates are actually quite good that you can get. So this is -- this is what they usually measured in bits per second per hertz per user. So with 10 users, you know, the average user is getting about maybe -- your medium user is getting maybe around .15 bits per second, a 1.5 bits per second per hertz which is actually a very good capacity. That means on like a 10 megahertz system you would be getting running an average capacity around 15 megahertz which is extremely good in a loaded cellular system. So, you know, power control has been very effective about getting even the minimum rate that users get can be managed quite well. So there's -- and it's a fact there's no user in a very bad interference environment -- interference condition. If the interference condition's bad, you can do -- you can do another thing about these power control algorithms there's a lot of mathematical theory about them. And then you can have systematic methods which try to improve by changing what's called the utility function of these users and improve their interference condition so that you can, you know, boost up the rates for the lower mobiles at the expense of mobiles at the higher end. >>: So that means coordination systems need to send signals to the femtocell base. >> Sundeep Rangan: That's right. >>: How much power should use. >> Sundeep Rangan: Exactly. >>: And then basically that coordination will help to improve the interference. >> Sundeep Rangan: That's right. I don't have time to go into it. But that's exactly it. But all that transfer of information is already part of the specs in these 3G and 4G cellular systems and work well for power control. All right. >>: So I mean this coordination does it need additional information for [inaudible] do you need the location [inaudible]. >> Sundeep Rangan: You don't typically need the location what you do is you measure -- do I have the slide on it? Oh, no. I have them -- but what you do is you -- the mobiles make reports so they listen to the other base stations and say this base station is 5 dB weaker than the base station to me and so you get these kind of path ratio and you send them, you keep on reporting them to your base station. So it measures -- it doesn't matter your geographical location is not what's relevant anyway, what matters is your actual relative location from the path loss. And those were always measured by the mobiles and they're constantly reporting that to the -- so once you start talking on your cell phone it will be constantly making those reports to the base station. That's how you can do the power control. And there's lots of other signals to do the power control signalling. I only have like a little while so I'm not going to go into it. But that's all there. But the problem is when you try to run these power control into a femtocell situation or just any kind of local [inaudible] they kind of fail. So here's the rate CDF, all right? So remember, so this is a CDF. So -- and this is the rate that you get. It's under some model, all right. But imagine that this is bits per second per hertz. So if this was a 10 megahertz then two would mean 20 megahertz, all right, so that would be a very high rate. So this is the situation. The blue curve, the situation all the links transmit a maximum power, all right. There's no power controls. It's the simplest thing. All right. And you'll see the many links, all right, again that, you know, 10, 15 percent getting virtually no rate. Right? Now, what happens when you try to do some kind of optimal power control so you say okay, I have some link in very bad power, I'm going to try to decrease the power on the other link to make a tradeoff so it can get less interference. And you can play around different games on how you do that tradeoff. But the reality is none of those curves really budge that very much. They budge it a little bit but really it's not helping out the situation. So there's still a lot that are suffering in a very bad interference condition. So the reason why it's failing is again that power control fundamentally works well when the interference is limited, all right? To illustrate that point, let's take a look here back to what, you know, information theorists call -- what the Gaussian interference channel. Let's take two links. All right. One transmitter and one receiver, another transmitter and another receiver. But they interfere with one another and this term alpha represents a kind of crossover game, all right. So this guy here will get some power P and noise N and interference alpha P. I'm still going to compare two things. One is reuse one. So we'll just assume here -- we won't even bother with the power control aspect. Let's assume that they both transmit at that time same time and they treat the other interference as noise, all right. That's when the dominant model for cellular systems. The other is I could manage orthogonalization, right? I could just say that well, I'm just going to split these two links. They'll each transmit on half the resources, the half time half frequency and then they'll put twice the power on that. So those are your two options. If you do this, you can look at the rates. And the rate you get will be a function of this crossover game, this alpha. So what you get is that if the alpha is very low, all right, that means that the weaks are a little interfering. The blue curve is a reuse one, the rate, and the green curve is the orthogonal. So let's look at the blue curve. When the interference is very low, obviously the reuse one does very well. But then as the interference decreases of course it goes down, the rate. At some point it would be better to switch over and just put them on to orthogonal, on to two orthogonal links. So this is the sort of high level conclusion of this very simple calculation. When the strong interference you want to orthogonalize it, when it's weak interference you want to do reuse one. And that makes sense of why cellular systems have worked well with reused ones because the interference tends to be low because of the proper cell selection. Now, actually historically reuse one wasn't always used. It was only used actually at the beginning of CDMA systems. 2G systems and that because they enable a concept that you could spread this interference across the whole band and get an averaging properties so you would see an average. But prior to that actually the dominant way was to using reuse like what you call reuse three, right, or something like that. And that was typically used in conjunction with sectorization. So you would sectorize your antenna like this and put your signal power on three different frequencies, all right? And now, for example, the green frequency here won't interference with the blue and the red. But of course what you're doing is you're limiting the amount of bandwidth, you're cutting the bandwidth down to a third. If you look at this here, does it actually improve the situation? And you could see here why it hasn't worked well in the cellular environment. Why you get actually an interference gain here, the amount of rate increase is fairly small. That's because you are sacrificing that one third of the bandwidth. All right. There's a small gain for people who have very bad interference, but the bulk of users actually suffer quite a bit, a large number of users. Let me just flip over. When you look at OFDMA, all right, what's happened here is that -- it's actually what you are seeing is in OFDM when you move to 4G systems they've actually seen a mixture of interference. So what's happened is that you have -- maybe I should just quickly -- OFDM what you have is that you allocate, you divide the signal into different times and frequency slots. Within a cell you can allocate users to disjoints parts of the spectrum. So for example these two different blue users are getting disjoint so they don't interference with one another. So they're orthogonalized. And then you can imagine in cell interferences tends to be high, so this is good. On the other hand out of cell interference you randomly hop the locations of the interference of these users within the cell, so they see kind of an averaging of the interference. So you're going to reuse one type of interference outside the cell. Does that -okay. Let me just flip because I haven't gone to any of the new part here. Let me just go to this. So if you look at sort of a, you know, historical perspective here on what's happened with reuse, early 2G systems used to be orthogonal. With the advent of CDMA, they went to reuse 1 because fundamentally that got a much higher efficiency because for the most part the interference was weak enough to -- was weak enough that you would get a benefit. So both the in cell and out of cell interference. For OFDMA systems, they got a kind of a best of both worlds. For the out of cell interference they could use reuse 1 and but for in cell interference they did an orthogonal. WiFi systems they didn't have a chance to talk about are really just generally orthogonal period. Because they tend to you have no condition of what the interference is and they just carry [inaudible] as soon as you detect any kind of transmission they shut down. So they tend to be orthogonal. The result, the WiFi systems tend to be simple and robust but you get a very poor spatial reuse, all right. But you have lots of spectrums so it doesn't matter. The problems with femtocells, all these kind of systems are based on certain design assumptions which, you know, which always are valid under a regularly planned network. But what happens in a femtocell situation? In a femtocell situation we saw here that reuse one does not work well in a small perjury but small but significant percentage because the assumptions break down. If we went to and all orthogonalized system then we'd have the problems that you'd have with WiFi which is very poor spatial reuse. So you need some kind of method, which is a hybrid. And that will motivate this last part of the talk here. So what LTE and 4G done in general is to kind of done a -- offered a way to do kind of a hybrid scheduling and what we call subband scheduling. In this case you can divide the bandwidth into a bunch of subbands which I've illustrated here. And the base agent can put different levels of power on different subbands. And now you can offer subbands at different types. You can have subbands which are -- they exclusively use one by one base station for orthogonal users. You could also offer base -- subbands which are used by both base stations. Now, as a result what happens, you can schedule users in different parts of the subbands. Mobiles that are not experiencing much interference can be scheduled into one of these reuse one subbands where, you know, they could actually benefit where mobiles at the end of the cell can then get the scheduled on one of these other subbands. Now, subband scheduling is not that valuable in the macrocellular environment because the macrocellular environment because it's already orthogonal in cell benefits almost -- gets almost everybody else you want to do in reuse 1. But this could be very valuable in the femtocell situation. In the femtocell situation what we want to do is some kind of -- you could think of distributed optimization to try to properly select the subbands and try to assign users to the cracked subbands. So you -- you could try to -- you would think of some kind of distributed optimization, not going to have a chance to go over this, would you use a utility based metric for this but then it's sending that metric in some kind of distributed way. Now, because the distributed here because well of course the selections are subbands in one base station will affect the subbands in this situation. When you run that algorithm, here's where you could start to see the gains. The blue line again was when all the mobiles were transmitting at the same power and the green line was transmitting with power people who got it a little better but you can get a much more significant gain here by doing this kind of power -- by doing some kind of subband scheduling. And that -- because what you can do of course if the fewer mobiles that are suffering very badly you can then go and assign them dedicated subband where they're not seeing any interference and get their interference condition off a lot better. All right. That was just a brief talk about what happens with femto to femtocells, let me just quickly talk about what happens in the femto to macro case. Now, the femto to macro case is -- creates a lot of problems. Logistically doing femto to macrocell coordination is difficult. For one part it's just a scalability issue when you want -- it's not really feasible for the femtocell large macrocell serving hundreds of mobiles and with potentially hundreds of femtocells in its thing to try to coordinate individual interference subbands with any of these parts. It's not going to say, you know, do one tiny little femtocell over here, I'm going to vacate this subband over you for, you know, which is going to affect all my subscribers here. There's also just, you know, a hearability issue that the signal -- there's a huge power disparity in the signals, so it's actually just very hard to -- or at least over the air to communicate any kind of signalling quickly between the femtocells and a short range communication and a large -- and a macrocellular environment. The second part is that even if the macrocell could hear you probably would never want to yield traffic. Macro capacity as we said is very expensive, it takes a lot of effort. The operator's never going to give away capacity of a macrocell to try to serve just some small femtocell. So ideally you really want technologies that operate under the radar of macro. Now, I wanted to say this because this is sort of the perspective in a way, a kind of dynamic version of cognitive spectrum sharing. Cognitive spectrum sharing you could think of as a slow time scale orthogonalization. When the primary units are occupied the spectrum all the short range communication's shut off. When it drops, you -- when it drops then the short range communication's detected and can come on. What we -- what we need for femtocells is something like that, but we'd actually like both operating simultaneously but the short range still running under the radar of the -- under the radar of the primary user which is the macrocell. You can do that trick doing subband scheduling. How can we do that? Let's take a look. Suppose we divide -- take the subband scheduling as we said. Imagine we divide it again and what we said reuse 1 subband and reuse 3 subband, all right. So reuse 1 subbands are typically serving mobiles that are close to the cells. So this color picture is -- imagine on the right is a blue bay station and the left is a green bay station. All right. So the power's kind of radiated here. The reuse one subband's both bay stations are transmitting and they're serving mobiles who are close to them so. The blue mobiles are being served by the blue base station, the green by the green base station. So obviously what can happen here is in the reuse one subband statistically there will be very few in the middle of the cell. So if your link is at the edge of the cell, you could kind of transmit here, you could imagine, a short range link in between these two subbands without affecting the mobiles at the edges -mobiles close to the cell. But statistically you should be far apart. This is coming at much lower power here. And anyway, you're not seeing much interference from this guy because it's very far away, and you're not going to have much in effect on the mobiles here. On the other hand, if you end up in the unfortunate situation where close to the base stations that's not bad either because the close base station will have reuse three subbands, and those subbands here will be serving mobiles at the edge -the edge of the cell, so you can run on one of those subbands this -- like on this one, the green base stations using so the blue base station's not even using it. So if you're close to it, then you just go ahead and transmit your short-range signal there. Does that picture make sense? Did I lose everyone on this picture. Or was it too many colors? [laughter]. Okay. They're two base stations. There's a green base station and a blue base station, all right. There's the blue mobiles being served by the blue base station, the green mobiles being served by the green base station. They are the squares. The red lines are your short range links, right, which are not getting served by either -- by the base stations. They just have to transmit without affecting any of this other communication. >>: [inaudible] basically the [inaudible] links what are the basically tricks you are using to basically make sure the basically the power for signals from base station doesn't overwhelm. >> Sundeep Rangan: So this part here I'm doing nothing. All I'm doing is just saying well, as long as the signal is weak enough. >>: Okay. >> Sundeep Rangan: So imagine that I am very far away from both the base stations. So then hopefully the gallon is weak enough I can transmit a little above that signal and then I'm not going to affect it. >>: [inaudible]. >> Sundeep Rangan: Nothing [inaudible] other than -- I just measure the receive power -- I measure the receive power and then I say if I look at how much power I need to transmit above that to get a decent rate. And if I'm far enough away from every other mobile, then go ahead and [inaudible] all right? And I can be far enough away -- okay. And I'll be far enough away because in this subband here I'll be -- those mobiles will be close to the base station. All right? So that's the -that's the idea. On the other hand, if you're close to the base station that's not so bad either because what you do is if you're close to the base station there are subsubbands which the cellular network will have content open for serving it's low edges cell mobiles. With its edges cell mobiles it will serve because this mobile wouldn't be scheduled here because if it was scheduled here we would see a lot of interference between the -- it would see a weak signal from the green base station and a -- you know, an equal power signal from the blue station. So you'll serve that cell here. This subband is one with a green base station only transmit when the blue base station is off. So you can trans-- if you land up close to the blue station you just run your short range link here, all right, because you're not even seeing any signal from the base station, so you won't affect it. And presumably the green mobile will be far away. Similarly if you are close to the blue base station you transmit here. So if you're close to the blue you transmit here, if you're close to the green, you transmit here, if you're close to neither, you transmit here. All right. So somewhere for you to transmit, wherever you go. All right. Now, a little part about science fiction here. You can even do better than this. Using something called interference cancellation. Interference cancellation is this. If it's very strong, actually, the interference you can first decode it, subtract it off, and then transmit underneath that. That opens up a couple more points on this color coded graph. If you're very close to -- if you're the edge of the cell here, imagine on the uplink, all right. In the uplink, this guy will be trend in very high powered or [inaudible] base station. So you can very clearly hear this guy's signal. You can then subtract it off and then transmit underneath his signal. All right. If you believe in science fiction you can do this, all right. On the other hand, if you are the edge of the cell, you can actually -- of course if you're the edge of the cell, you can always transmit here because there's no signal. But you can also transmit when you're -- when this base station's signal is very strong. Because you will hear on the downlink you'll see a very strong signal from the base station, so you can then decode that signal and subtract it off, all right. Now, of course for interference cancellation to work, you have to have -- this has to be a decodable signal. It means that the rate has to be relatively low so you can decode it. But any way serving edge of cell mobiles. So any way, the rate tends to be very low. So -- which is -- which will be decodable. So you can just make this so that these are -- so you'll tend to be able to decode the data here. When you add all of that up -- yes? >>: Interference cancellation is basically the hot [inaudible]. >> Sundeep Rangan: Yes. >>: [inaudible] what are the assumptions basically in the interference cancellation and how much of the assumption can we implement in the real world? >> Sundeep Rangan: Okay. So the interference cancellation just really -actually it's already -- there are Qualcomm when I was there is already producing chips, interference cancellation. Interference avoidance is another -- but interference cancellation is you just -- the assumptions of course is that you have to be able to detect the assignment from the base station in this case. So that part you just have to put the assignment at some publicly decodable part. It has to be at a rate that's low enough that you could decode first. So, you know, you're trying to do a very high rate to that user, you know, you'll need a very good channel quality and it will be harder to decode, all right? So again, you have to lower the rate. That's easy. And then you need hardware to actually do this. The hardware is a bit of a problem, especially -- it's really -- you know, I've been through the basic part of this part and you need really the bottleneck is actually the memory to get that. That's because in a cellular system you have this thing called hybrid RQ where you have to remember codes and then try to decode them. But the memory's actually becoming, you know, reasonable as well. You know, at least for these cellular systems. So that's -- it's entirely feasible to do this. >>: So [inaudible]. >> Sundeep Rangan: Yeah. >>: [inaudible] so basically for interference cancellation the receiver will need to try to be able to decode two signals, one is a [inaudible] signals. >> Sundeep Rangan: Yes. >>: Which is basically from far away [inaudible] base stations then the basically [inaudible]. >> Sundeep Rangan: Yes. >>: Right? >> Sundeep Rangan: Right. >>: So basically you try to do decoding twice. >> Sundeep Rangan: Exactly. >>: Right? I mean, basically first strip of -- from the one that ->> Sundeep Rangan: Yes. >>: To the lower ones. >> Sundeep Rangan: Exactly. Okay. So when you add this all, if you believe this is a very preliminary simulation here. When you add this up, you get this potentially huge amount of capacity [inaudible] 10 megahertz system. The medium user can get what, maybe 25 megabits for free off this. There's some minimal assumption here, some minimal affect on the macro. So like 25 megabits for free significant underneath the radar of the [inaudible] and with interference cancellation that boosts up to around like 35 megabits. So it's actually -- so this, I think interesting just from all sorts of people looking in cognitive spectrum about a different way of thinking about how to share spectrum that's maybe not so with just a simple either or kind of situation that there's a lot of capacity. Let me just tell you -- I can talk about this later. Maybe I'll just wrap it up right now. We'll talk about the test bed plans that I have for poly and try to build a wireless ->>: [inaudible]. >> Sundeep Rangan: Okay. [applause]. >> Sundeep Rangan: Sorry to keep you here.