All. Right we're gonna have a lot of fun today, we're gonna have a lot of fun in this. This. Session, we're, going to keep it very engaging. And. Which means that you, know you need to, step. Up and ask questions and, challenge the speakers challenge. Me the. Way we're gonna run this is, in. The first 10 minutes I'll speak, and. I'll tell you a little bit about my perspective, on this it's, about ten years since I started working on this in this, area, and, I want to sort of share that with you a few, of you may have heard this before because I actually had a little workshop, last. Week and I said the same thing so for. You guys just focus, on your work for now but. Anyway. So ten minutes I'll speak and then we'll give twenty minutes to each speaker, that. Takes and there are four speakers that takes us up to one hour 30 minutes but. We will not end there the. Way everything you set out we actually do have a half our budget and, buffer and, I'm, gonna use that to. Have. A discussion with all, the speakers I'll. Ask them some questions I, expect you, should ask them some questions and. We'll, make it interactive there. Are five seats here or, six what is it yeah, so. There's one extra in, so, we have a surprise. Guest. You. Know who you are you're here already okay. With. That let me begin. All. Right this session for those of you who, is. On intelligent edge so if like. They say in the plains you know if you're not the right flight you probably want to go. Okay. That joke didn't work sorry, okay. So. I'm, gonna say a few things about my, own personal journey on this so, before. I say this has been hammered, into you I think by. Now but I want, to take this time. To do it as well, Microsoft. Azure is, large, it's very, very big it's very complex. You. Know serving, as these numbers sort, of indicate, there. Are so, many moving parts as the network, is massive. 33,000, plus miles of lead fiber across. Our lands more than hundred thousand miles of fiber inside, our data centers. Millions. Of components millions, of servers all, interconnected. Packets, flowing from one point to another point, it's. It's an engineering, marvel, in some sense it's planet-scale, it's huge, and as. You think about research, you can, start to start to see that you, know we, need discipline, a lot of discipline, in designing such systems and we. Have to be very careful, because if anybody's. Gonna take dependency, on us on the, cloud we better damn well stay, up all the time right, that's why people work very hard and, they. Worry, about anything, breaking, at any given point anywhere, such. A large, large, large. System. In. Terms of in terms of the economics, of it you also heard Marc talk, a little bit about it I had these numbers to say I want to share them with you as well NFI. Eighteen twenty, three billion dollars, revenue that's, a significant, amount and it's going fast up eighty, nine percent growth but I have these numbers are from my f17, I haven't wider. Updated, this morning but I couldn't but you can sort of see the numbers how, fast, the growth is happening we. Can't build fast enough that's. The deal there's so many customers. We, can't build fast, enough in fact, we have to slow down to, prevent. To to to, avoid. Over-promising. But, from a research that actually makes an interesting problem to I mean if the cloud is could. Be statistically, multiplexed you, know we could actually use this service again so there's, a lot of researchers at Microsoft, in my group and close. Eye groups that that do that and so, you can also see from this number 90% of the biggest largest companies in the world depend on it so it's, all good so that sort of sets the stage for things, all, right so, how did this thing start for. Me, in. 2000. Have. The right idea in, 2008. I invited, satya from CMU many of you might know him too. To. Come to microcell research in building 99 and, we. Had a one day brainstorming. Session on, what. Does it mean to what is going to happen after cloud things. Change - in fact I would challenge, you to think like that - 20 years from now what is the world going to look like right so that's kind of sort of what we were talking. About and. Then our other discussion, came. Out this page paper this. Is the. Paper that people saw that the site. As cloudlets, and we. Published this paper initial. Thoughts you, look back at it you sort of think okay well you know we could do things differently then. This is the first article, that came out and December 13 2009, which. Says how the cloudlet, will beat the cloud this is the edge so think of it as the edge the, intelligent, edge and, if you look at the citation count which I just did while, ago you, can see a paper is very heavily cited so a lot of people have read it or working, on it in a in a publishing on it now.
In. The faculty, summit of 2009. I did. A demo of edge. Networking, and the demo was the actual what is now the Skype translator we. Literally, I mean. I was sitting here Otto is with me and we offloaded. This. Translation, capability, to the edge on, a phone and then he was talking in Spanish I was talking in English and we were doing this in real time and. Then we published the paper with, the programming model for that of how, programmers. Will, do that and that, shoe has, been cited quite a bit, since. Then, okay. So. Then I had this opportunity we have in Microsoft research we have this, thing called the DTR or disruptive, technology review where researchers. Can go every. Year happens, every year now starting to happen every two years but, can sort of go and talk, to the CEO, and the entire SLT they, come into the, building ninety nine and. Tell, them about what, trends, or what things we think can either, disrupt, us or we. Can do to disrupt the industry and so. In that I actually presented this I presented. This if we look at this picture you will sort of see this cause Maui node that's the paper that was referring to but really this is kind of the edge right and this, is the cloud and, one of the things I said that think, the slide is from, that time was. That here you have this wireless things you have access points you've got the switches. And here right next to the switch is your edge note that. Was the imago you know it and then you, know we talked about it but, I made, a mistake, Steve Ballmer who was a CEO asked, me okay, Victor what. Do you think we should do differently today, right. He didn't ask me about the future he asked me what do you want me to do today and that, was a question I did not expect the shame on me now I look back at my life and think that was shame but think, about it as researchers, we always focused right, you're always saying oh we can do this we do, this well, he was here to see you asking me what do you want me to change right now and, I, stumbled, on when I told, him speech. Speech, recognition I think I said the speak translation, or something and. There, was an EVP there and he didn't buy into it he said not, needed and, in some sense he's right and when we have all this stuff going in the cloud and works but, so that was sort of didn't really work so I went back and. I you, know I kept working on this stuff I really believed it and then I had another chance in 2014. Actually mm that was for 2013, but we did it in 2014, I don't, know the chance in front of the SLV but this time I took the problem on differently I wanted. To make, the case that look. In the cloud we sell storage, and we sell computes. I want. To sell latency, ok, latency is the hard hard, thing because you don't have control many of it you don't control the internet so. Then logically. You will immediately says well one thing is to make many. Many data, centers around the world the, other and that's not you, know financially, viable so the other is hey let's put lots of edges around the world right so that was thing and I showed in this video which I want to show you I, do. Have a point after all this so let's not spare me for a second, so. I wanted to watch this video carefully and tell me what you see, so.
There Are seven people walking, there's a camera that's taking this the original video right now, we're gonna do face recognition on it, with. Two to five millisecond, latency on. The edge and. You should see that. And. Then we can run the same program this, is four years ago right so things, are better now with, 7585, and the reason we used 7585. Was because. LTE. Rule. Of thumb is about 70, millisecond, latency what, did you see different, in the two can anybody. People have seen it don't please say whatever did, anybody see anything different in this. Come. On. It. Wasn't. Okay. That's Ana's Aysen. Anybody. Else, would. You like me to play the game. Yeah. Yeah. That's, true - I'll. Do it one more time and one. Is it worse so there's the original original video then, there is a video with face recognition five, minutes five veneer, floating, 5 milliseconds away time and then, there is. 77. 5 milliseconds away afterwards. So. That's the first one and. Now. Here's the second one with, the LTE Layton sees. The. Do you see this before that's. Exactly, right there, were seven, people seven, were recognized, in the second case there were only four that's. A serious. Problem that's a functionality, problem, imagine. That something. Like that somebody, else does and you're. You're actually losing serious. Functionality, there you can't do it the system can't do it right it's not fast enough so this is what I showed I actually, showed a bunch of other things - well I'll show you another one which. Exceed. Mika-chan does this is gaming fast action gaming that. Was happening and. So. Eduardo you know he's playing this game he's a he's, a top-notch. Gamer and, he. Plays the games with five second 30. Millisecond, latency and. 40. Millisecond, latency now this particular video all you're seeing is I just point out because we, don't, have time here is really, when he clicks there's lag right, between the click and what he sees so now if you transmit, this, to a surface, more, serious game where he's fighting somebody and it's, shooting what. You see is that as, we increase the latency he. Goes from killing everybody and winning the game to. A point where, he gets, shot all the time because, he reacts but.
His Reaction time is slow in, the game is still playing and boom. He loses the game so, that's. Not gonna work so if you do say cloud gaming you know it's not gonna work unless you match the latency so I'm gonna just move forward. So. So, I made this prediction after that in the DTR I said, the world is now moving in this direction okay so you got this mega clouds you, got to have a disaggregated, cloud you know not going to be able to sustain mega clouds for various various reasons and it. Was sort of the, prediction was raised by 2020, everything, is going to look very different. Alright, and I was actually when, I'd made this talk I had actually gone through many examples, of what are the different advantages of edges right. You can read this I'm not gonna belabor the point because you can hear more from people here but, this was not just a prediction made a sewer systems built showed, this, was happening so. What happened after that meeting is this. Is just before Southie, became the CEO Steve. Stop Ballmer asked me how much is gonna cost in Satya, who, was in the audience said three hundred eighty million dollars for us to go do this and, so, and Steve said do. It do it, I'm. Like shocked i won. Well. Not, really it's one thing to get senior management, to agree it's another thing to get the people who actually do the stuff to agree and that's, where everything gets stopped so. What. In the meantime the, world was kind of evolving, - I mean Satya had gone and created his own you, know if you go he, has done tremendous work since then he was a he's a true believer there and then, I've seen the world sort of moving all the different directions as well a lot, of like this. Conference was created I was part of that and, before consortium, was created by, the way I, hate. The word fog hey and I also think they ripped off many. Of the things we did and just call it fog but that's okay and I actually have proof for that but that's not so. Then but we still I was still not happy, because where's, the money, where.
Is The money we needed a killer app so. Then we had this observation, there, was a lecture in sabbatical and the solution happened that, there are cameras, everywhere. There are millions, of cameras, everywhere. They're, streaming video, nobody's. Looking at that computer. Vision and AI were becoming big systems. Were good so, then we started working for four years we have worked on video, analytics trying, to show and yes. I still believe there's a lot of money here and I still believe this. Is the edge story from my perspective, and. We did some good stuff here we, actually won two national awards, because. The work we did here, then. Something interesting happened, I. Always. Worked I was using this work yeah I mean like Microsoft, talking to all of my colleagues and, then in. December, of. 2016. This guy he's. A VC if you don't know him he's, a well-known we see from, Silicon Valley he gave a talk entitled, the, death of cloud, in. That talk, he. He basically made the point for intelligent. That she talked about AI he, talked about you. Know. Doing. The latency thing pretty much very. Very similar to what we had been sort of talking about here but. It was outside it wasn't the VC he was you know he's a he, funds companies, and things of that nature, boom. It became a it, became a big thing in. The meantime well, you know I have great respect for our CEO and the reason I have a lot of respect for sat there is well. First, he's very he has empathy and I like that but another, thing he's a smart guy he's, his he has. He. Has his finger on the pulse of what is happening in the world and he watches it he observes it he talks, to it and then he makes decisions right, so, he's. Along he knows we're doing this but then he starts, to see this and then he. Goes out there and says, intelligent. Edge we are going we're all in it Microsoft, is all into that it's. Not just the cloud stories intelligent, history now, things, start to move in a really really big way so hey this. Is what he says in Bill, 2017. But he, had started saying this already as you can imagine. In. The meantime a, lot of research reports start to come up, edges. That edge is gonna be, the thing and this is just one example of, one of those reports there, are many who. Came. Over about how the edge will eat the cloud I actually, don't believe that at all and we will bring that up into the panel and then, of, course you start to see news articles, you start to see people getting excited there's, so many I just. Some samples here just to say that, you know all, these, it. Was becoming a big thing all. Right then, another big thing happened, so, Microsoft announces. That we're. Gonna invest five billion, dollars, in IOT, and in. The text, it. Talks about the, intelligent. Edge as one, of the key components to. The five billion dollar investment, now. Things, are hot really. Hot you've got a hundred, and ten billion, dollar company, saying. That we're gonna start investing, in this, that's. Significant. Now. People, are starting to move pretty hard you're gonna start hearing Arjumand, is one of our speakers and he's going to fill in a lot, of the gaps that I'm leaving here but. I'm just taking we're talking about it from my perspective of how exciting things are getting for me, then. A couple days ago I am the p-switch I was a beast year of this symposium on edge computing, I'm, sitting there and, listening to the papers you know if we had all these paper discussion, to figure out which papers to accept I. Start. To see papers, in all these different areas but. Age-related. As. Related papers this. Is really exciting now we've got businesses. Starting, to say this, is big, you got the VC community saying, we. Are ready to invest you. Got researchers. Pointing, out all the different issues which is going to keep them alive working. For several years and these provide these problems. And their solutions are, quite innovative if you think about it you.
Have Something now it is not no longer a fad there's nothing that this is something very concrete because these problems. Will get solved for example, one of the problems that still exists is availability in. An SLA in clouds we have SL, A's so, if you depend on something you say it'll be there when you introduce edge what's the SLA what's the evil availability, a lot, of papers on that yeah. And, since, it's early work they were all these the whole so we didn't accept all of them but you, can sort of start to see this you, saw things of like management when you have imagined, that you've got not just one is now you've got tens. Of thousands of s way more than the cloud how are you gonna manage it who's gonna own the edge is it gonna be like the. Companies, like Microsoft, or is it going to be sort of a federated model where somebody else on the edge we don't know so, as you think about all this stuff you think what, about privacy what about security in Flour. We can offer you some level of security because we kept everything in lock and key now, the edge is sitting on your premise somebody, can come take out the disk copy it put it back you know who, knows what happened right real. Serious problems come on so. So. My final thoughts here are the, there is a paradigm shift that is happening and it, is going to be very big I predict I, just. Looked at some report and it says this year is the year of AI or an edge I didn't, focus too much on the intelligent part of it I'll leave the speakers, to do and, with, that we start this thank you and then we will start the thing, by having four speakers I'm really excited. To have them is mom here. Yeah. Come, on are we going to speak in this order so, let me introduce you, to, monk. He, is many. Of you might know, him he was actually the director and investor, of the. Edge lab in, Princeton. So well thank you victor very much what a pleasure to be here thanks for the invitation. And, as. Victor said that. My. Day job is the Dean of College of Engineering, at Purdue now since. A year ago but. A mine nighttime. And vacation, and weekend is. Research. In networking and continue, to do that part, of that is still. A visiting, appointment, to direct edge lab at Princeton and the other is. Participating. In the open, fall consortium, a victor, now we have debates about the use of english words like fog or edge and, i'm gonna just brush, that, aside for, the next 15 minutes here since, the last five minutes trying to get Mac and the, Microsoft, to work together I'll try to be brief here and. The main message is a network. Decomposition. And what is that. Well. The job I like the CEOs, job you've got a great one and. Victor. Is to sort, of decide who does what at, what time scale, and how to glue them back together so. It's not just allocating. The, resources, but. Also allocating. The functions, and. How. Each, function is carried out is less or becomes, for. The. Decomposition. Itself. Now there's the vertical decomposition. Layering. As. A fundamental. Concept in networking, to modular. Eyes a complex, interacting. System, and layering, candy. Mathematically. Understood. As, optimization. To conversation, both reverse engineering, and also a design, synthesis. For engineering, and for. Example Kumar, was. One who gave me and, many in the community a lot of inspiration, back this is 15, plus, years ago also on. The. Notion of architecture. And thinking, about layering that way now. This, subject, that Victor is putting together. Panel. This morning on intelligent. Edge now, I think of the cloud and edge slash fog as. A, horizontal. Decomposition. And. All the architectures. In some sense are there, to, be. A little, philosophical if. You will and. Then to support a range of practical applications. For, example in, 1948. And, the Shannon source channel, separation, is an architectural. Principle. And very. Crisply. Stated, mathematically as. Well and that. Principle. In. Architecture. Further led to the, applications, of digital communications. And when. Mr. Bob, Kahn did the tcp/ip, published. In 74. And that, is an architectural. Design. And. That led to the vertical. Applications in. The, internet and web and, the edge and fog. My. Biggest hope is that it will continue, to, be worked, on by the, community. From industry, and academia, around, the world and arise. These, architectural. Foundations, for, applications.
Such As IOT, and v g or, dispersive. A is. So. I'm gonna a brief abbreviate. My, personal, journey much shorter, here. Back. About the decade ago, started. A Princeton edge lab back, then I got, to interesting, question one is one large fortune 500 company told me the, word edge has been not, Microsoft, has, been trademarked. So I have. To stop, using the word edge lab and, the, Princeton lawyer wrote something back and, then never heard anything from them again the other is that people told me a moment you know well, you know I like, you and stuff but you're, on the wrong side you know everything, will move away from the, edge edge would be dumb have. You heard of the word cloud and. Distributing. Functions, weathers computation, storage. Or computing, or control, to the edge and. Then back, to three years ago, was. The founding, member, of the open fog consortium. Now. One could ask why. Do you, not like the word edge anymore, after. A few years I say well not really I don't particular. Care, about English. Word and. I, know that for example Victor. Thinks that fog is a very vague foggy word, and why don't we just keep using the word edge. But. You know if. There is an industry. Including. Microsoft, Cisco Intel, and quite, a few others want to get together use a term I say that's fine with me too and. Decomposing. Functions, along, this, c-to-t. Continuum. Cloud to things continuum. So. This is sort of a, wrong. Picture to show but let me first tell you what is. Perhaps. Useful about, it first. Now on the one side you've got the cloud on the other side you got that, looks like a fog or mist or whatever you call it now, the. Interesting point is that you need both. Now. Peter Levine's. Provocative. Essay. To say the death of cloud you know might be exaggerating. To make a point I, don't. See the death of cloud I see the rise of the, edge and fog and depending. What you want to do, what. Belongs to Caesar goes, to Caesar what belongs to God goes to God and. Now. This picture is however also, wrong because it. Is not a dichotomy, it is a continuum, there is no digital sort. Of boundary to say yeah is, a. Choice. Between, either. And or, and. Furthermore. The. Pendulum, swings back and forth the, time scale is about what 15 20 years or so. So. Some. Argue that there is nothing new Under the Sun and then merely. Swinging, the pendulum every two decades or so whether. There's NSF, or industry. Or PhD. Thesis, and perhaps. Is not that cynical I think every swing brings new ideas and, driven, by new, context. - but it, is true that the. Rise of the cloud and. Then, now there's the rise. Of the age and fog then a big, believer just, like Victor, myself, but I won't be surprised, that 10-15, years later at Microsoft. Faculty summit whatever. Twenty thirty four, people. Would say oh you, know this edge thing is just so like, 2028. We. Should go centralization. We should look, at economy of scale and somebody in the audience will say I regularly remember, when I was young there's something called cloud, so. This pendulum will keep swinging, and, and that is fine. There. Are many topics as Victor is saying one could talk about in in, edge intelligent. Edge or fog. Including. I think the economic. Side of it is very important, how do you incentivize. The. Security, privacy is important. But. Let me just focus next, ten minutes. Particularly. On a dispersive. AI as. One of the in. My mind three main pillars self-application, verticals. So. Here you again, want to leverage the, fact that you got latency, on your side, latency. In two ways first. You, can do edge analytics. Or stream mining, with the, life. Streaming. Of data coming to you and you. Can understand. It right here right now the. Other is you dare to put in some feedback control loop because. You know that latency, and perhaps more importantly, jitter is much. More under control is small and. Almost, deterministic, latency, so you can run control loops in, cyber-physical, systems. Now. You got to decompose the, machine learning computation. And control into. Disparate. Components. And then they will be working. Either, collectively. Or, jointly. And. You, can think about pre. Positioning and, so on. So. This is about. A year ago we did some demo, of one. Particular. Platform. That we did at Princeton. Think. Of this as a MapReduce, for. Ads. Or fog, and what. You want to do is to. Look. At opportunities to, frame. How. Would you decompose. Machine. Learning computations. In, order to map to this new substrate, okay. So I'm going to fly through some slides. Here but the. Typical chart, you want to see this is just a toy example, last. Year is to say well, between. The one extreme and completely, on the device the other extreme, and completely. On the cloud what. Happens in between, well. There's a trade-off between what's. Done locally and, what is coordinated. Community communicated. Globally. You want the summation, to be smaller. Than. Either end, and.
That's. A little comparison, between what, we did and spark. The. Unique feature really, is that you've got an edge or fog. To. Size. On the, challenging. Side. There. Is heterogeneity, unlike. The cloud, you. Don't have racks. Of servers that, run. On, the same OS, and. Similar. Hardware capabilities. You've. Got, variability. Even volatility. Because. Unlike. A rack of servers these, little, things they, move perhaps. And, they may not always be available, and. You got constraints, bandwidth. For sure and maybe even battery. Constraints. But. On the other hand you also got the, nice, things such. As proximity. Physical. Proximity. To, sensors, and actuators. So. The challenge is to take care of heterogeneity. Variability. And constraints. And take, advantage of, the. Proximity, now. How would you decompose, a job, you, know whether is you know counting, words, in a file or slightly, more interesting, let's say. Run. Run of least squares for example, all right one, of the simplest things you can think of if you run a large-scale. Least-squares, as embedded, in the, loop of some, particular. Optimization. The for machine learning you, know how to do that by. Decomposing, the job into, parallel computation, on a rack of server very very well but, how would you, decompose. That job. Over. A very different substrate, the. Ad Fox substrate, so. Here are two tort examples, to illustrate, in. This notion of decomposing, the, computation, over this kind of substrate. One. Is about junk camera working with, tension, surely, in, my lab. So. Imagine, now in this particular application, topic. You've got, the. Outdoor say. Sports event like, the. FIFA. World Cup soccer and. People. Have thought about using. Cameras. Over cross wires for, example, but. If you use Jones. To. Capture and to transmit. What, would happen right. So. Here is sort. Of one way to look at the choices are of architecture. On. The one hand you can make, everything locally, done. But. Then you lose the. Collaboration. You, may miss, some of the important scenes in the coverage and you may have to reduce the. Video. Transmission, right out of those drones, on. The other hand you have a fully centralized, version, of it and. You. Have to make decisions. Essentially. Now. This. Cloud could be the, azure cloud, way, out there or it could be even a big, server on the ground that. Big server to all the little Jones feels. Like the cloud it's. Just like you. Know a little. Rock to us may, feel like a mountain to the ends so. But. You're, gonna suffer, the round-trip time that means the. Ability to coordinate in real time in, response to game dynamics will, be limited so here. Is one, particular decomposition. I'm not arguing this, is the, best the decomposition, even, for this, application, but, you see that the Jones form laid a fog net were among themselves they, would take care of the. Following functions, trajectory, planning and, collision, avoidance, and then, there is a little central, coordinator. For. Next target location, optimization. And deciding. Which Jones should be doing what so you've got eight Jones let's say up in the air which. Of those should, help you to. Cover. The. Scenes and which should, help you to relay. The, video streams so. That will be done. With. Coordination. So there is still, controller. Up. There and. Then you've, got the sensors on, the boss and the players, and, then you got the say eight whatever. Number of Jones up in the air, so. The Jones now they, got to decide, to cover they got to decide to transmit, and you can do other fancy things 3d, rendering, and so on if you want to do that in real time, and. There is actually a trade-off between. Maximizing. Coverage, and. Maximizing. Throughput, and you, only got that, many Jones, so for, each Jeong, should. That, Joan, help with coverage, or should that help with relaying, and. Enhancing, throughput, it. Turns out you can run an integer programming. Joint, optimization. And. There are some heuristics, will allow you to solve it very fast even though, not truly optimally. Just. A little chart, to show what we did I, guess. Illegally. On a field. On Sunday. Mornings, near, Princeton I shouldn't. Say that. Probably. Legally are flying. The drones and. You. Look at the trade-off over there.
In This particular case it, turns out that you've got a Jones fall, of them to focus on coverage fall of them to focus on streaming, is the best now which is covering, which is streaming, that, depends on the real-time situation, and. That. Is actually coordinated. By, the cloud, if you will so, the message here is, not about the application, per se we can go into that detail on another, day but. To say if you want to coordinate. Group. Of Jones working. With each other certain. Functions, can be done by, themselves, on the Jones and others. Should. Be done on the, ground with a little server. So. The second quick, illustrative. Example, is. Something. That just we were doing it last. Month it's still ongoing is, some preliminary results here to share. It's. About a RvR and particularly. The. Technical. Problem. Of object, placement so, that they. Don't block important. Things and don't, confuse the, viewer and. We. Want to personalize, the, placement. Challenge. The, question is where. Do you learn about the personal preferences, and, where does the learning happen. By. The way that is actually the Princeton, edge lab outside. In. The corridor over, there and those are the Holograms, in that. Experiment. So. Again the, main message is that there's a variety of choices, in, network decomposition. Here. During. The policy, learning, phase there, is a user, there's an agent, and the. Agent, will. Try to mimic the user and learn, what. Is the personal, preference, of overlapping. And foreground/background, placement. Of the holograms and then, there is secondly. A policy, execution, phase after. You've, gone, through the learning phase, and, in. Both phases you can either have a little server there or you can have the end device. In this case the, they. Are set. To. Do the computation, and. What. We found in this particular case is a trade-off whereby. The, policy, is. Learned. With. The computation primarily. On, the, little server, next. To these AR devices. But, then in execution. The, parameter, choices, of the, learned policy, is done in real-time and. That's not against, surprising. Trade-off, but, it's one that sort. Of worked and. This. Is. A. Quick. Video, one. Minute I'll probably skip in a few, seconds, on the left hand side will, be is, the, user on the right hand side is the agent, this is during the learning phase you, want to learn the policy, before you execute. The policy, in real time and it. Shows how, the, right hand side agent, is learning, quite, efficiently, based. On what the user is doing to. Start, it out if you still remember 15. Seconds, ago way, off you know whether, the. Hologram. Should be placed in, covering.
The Real object, or behind the real object, where it should be but. Within. 10-15, iterations, I learned very fast. Let. Me just skip, to. The. Next, slide, there. So, one way to visualize that a learning process is after. A few demos, basically, all. Those little dots, are. The, decisions. By the agent, learning, the user preference, and personalized, placement, and that, they all aggregate, to the correct, place now. What is interesting is that if. You. Don't. Use state space reduction. Then. Even, after many more rounds, used. To are. All. Over the places on the left hand side is you. Do use state space reduction, right hand side is you do use it and. You can clearly see the benefits, now, the state space reduction, is carried, out not on the, AR themselves, but on the, on. The. Little server so, again the main message. There is not about, AR placement. Personalization. Is that there's, a choice of where, do you carry, out the learning and. One. Possibility, is let, the policy stay, in. The cloud if you will and let, the, actual. Parameter, and execution, stay on the end devices, so. The overall question really is, where. As. In, the horizontal decomposition. Where, do you place, a function, and, we have him got to the point, of where you place the data and, what. Is the interface, among, these, modules. Over there so. My last slide is just to highlight what is perhaps obvious, on, why, is it even interesting, to, explore, that. Today. And I think of the acronym scale. I learned. That if you want to write a winning, NSF, proposal you have to come up with good acronym. Although. This was refuted, abaya, by by jim crow see last, night to me that it that's not, necessary. Condition, but I guess, it doesn't hurt so s is, the potential, advantage, on security, the less distance. It traverses more, privacy, is, potentially. Protected, one could argue the other way to say well these little things trusting, them with more, and more functions, but they are easier, to hack into, and their update. Is not done properly. Cognition. Is because. You are much closer to the end-user an ambience, you. Have a better understanding, of user, centric, objective. Agility. Just refers to the fact that generally. Speaking you. Go into the cloud the, core network of LTE and so on you go through 3gpp. And it takes forever to. Update. And change, whereas. If you rely on that end devices. The. Agile, development process is, much, faster. Latency. Were mentioned that's the number one reason as Victor also highlighted, that people think about why do I even care and need, anything. Other than the cloud and, efficiency. Is actually, interesting, there's, the D, to D efficiency. Ok, device talking to device but, then there's also the device. For, device before de efficiency. That, because, you have proximal. Distance. You, can Luber eyes the idle, resources. Connectivity. Computation, and so on there. Are in physical proximity to, each other, so, SCAL. Ye these are the reasons, I think of why networking composition. Horizontally. Closer, to, the users, add. Fog. Is. Becoming. Ever, so, more interesting, thanks. And. Ask the question and I'll answer. I. Will, do it. So. Great. Pleasure to introduce Amaka, shore Ramachandran, he, actually goes by the name kishore. Anyway. Okay, so anyway so, he got his PhD from University of wisconsin-madison, he. Is now a professor, at Georgia, Tech his. Interests, are in parallel. Distributive systems and. As. You can see he's been working, on this space for a while, he's quite passionate. About. Providing. Connectivity, and, services, in the rural areas I actually was. There in one of the NSF. Workshops. You had actually organized with Aneesh and. Quite. Fascinating, with that please take it away well, I mean there's a debate. That. That. Vikram started victor, started with edge.
Versus Frog but, let me ask you a question Oh. Victor, is real first name any. Of you know. Among. What is the what is the name you know. Part. Of me Paramveer. Is a real, first name and and. And, it, means supreme. Hero. And. You should ask him during the panel during. The panel part ask him where he got the new holy got the name victor. So. So. I think moon, and. And. Victor, both of them gave a nice. Segue. To what I'm going to talk about and. My title the title of my talk is, as. Making. Elevating, the edge to be a peer of the cloth so that's the. Field. That I'm going to talk about and the. I. Mean of, course I won't be standing here but for all. My students who've done all. The work there. Sort. Of highlight, a little bit in, describing, what. We're doing so my. Thing IOT, and the. Scale of applications. That run on the IOT all these are things, that you all are familiar with and and, the key, thing that among, and Victor pointed out is that up, future applications, on. The IOT infrastructure, is going to have this characteristic of sensing, processing. And actuating, and which, means that levy, becomes, very critical in going. From sensing, to to. Actuation, and cloud. Computing as we know it is great for throughput oriented application, but not latency, sensitive applications there, are also other considerations, regulatory. Considerations and, backhaul bandwidth and so on all, of these may, argue, for moving. The computations. Away. From the, cloud and and that's where fog. Edge computing, comes. In and and. There is a moon. Pointed out it's, really a hierarchy, of devices. All the way from you. Know the client devices to routers. To service. All the way to the cloud right so that's the continuum. That exists, in, this in this world and. Now. What. I want to. Emphasize. Is the fact that today edge. Computing. As we know it it's, a slave of the cloud and and. You know there are lots of very, interesting. Platforms. That have been designed aardman is going to talk about IOT as, your edge cisco. Has its own iox. Plat platform, but, all of these require. That. The coordination, is coming from the clock and and, and and. There's a great example that that. Move already pointed out in terms of drones coordinating. Is one another so that's the idea behind, warning. To sort of elevate, the. The edge to be a peer of the cloud and today, actually we are beholden, to the cloud because if I want to send a text. Message to husband, I have. To go to the cloud and come back to him right and so. That just doesn't. Seem. Right to us and so. Vision for the future is that the. Edge is going to be becoming a peer of the cloud and this, prior art that, Victor already pointed out cloudlets. And, and, and and Maui and and, all of them are, trying. To do the following which is sort, of take computations. From, the as a client devices and launch it onto onto the onto, the server that is nearby but in the limit what we want to do is make, the edge autonomous, even if it is disconnected, from the cloud and I can give lots of reasons why this. Might be a good thing because. The interacting, entities if, you think self, connected, you know connected. Vehicles applications, and so on there. Is lot of things that need to happen horizontally. Among the, edge, nodes and and. And even, more importantly. If you think about disaster. Recovery and things like that we may not even have connect connection to the cloud and in that case and, the drone example, of the great one right so if if. I have the ability to, connect, to even. If I have this connections, among them if, I have the ability to connect to somebody I can actually communicate and, then eventually my, information, will, get passed on to the others and and, we, can have a happy situation so these are motivating. Reasons why you, may want to elevate. The. Edge. To be a peer of the cloud and and. There are lots of challenges making, this happen, first. Of all as a cloud as we, know it today is so, powerful, mainly because of all the infrastructure, that have gone into it right there.
Is You know programming, model, storage abstractions, pop, cell systems and so on that, make it a powerful. Development. Engine. For, rapidly. Prototyping, applications, and, and, now we have geo, distributed. Edge. Nodes and we, want to provide, the same sort of ability for, rapidly, prototyping, applications and, what, that means is that we, need powerful, frameworks. Akin to the cloud to, be available at the edge and. Yesterday. I, will. Bring Greenberg, was emphasizing. The. Need for, making. Sure that recovery. And is. Important, and which means that when you have geo distributed, data then. You have to worry about replication, and Judis in fashion so that you can avoid you. Know correlated. Failures that might happen in a particular neighborhood and so, those are important, and when you do that then, there is the latency consistent, latency, and consistency. Trade-offs. That comes about right so. Those are things that you have to worry about and. And, and, then similar to the cloud be warned that is the same sort, of elasticity, of. Deploying. Application. Components, rapidly, and the, fact that you might have the same IOT, infrastructure. That, is simultaneously being, used by multiple applications. Makes me it means that you have to have worried worried about multi-tenancy. And all, of these are issues that are. Challenging. And in the remaining. Time. What I'm going to do is you, know lots of people are working on interesting, areas, related. To these topics, and I'm going to give you a little bit of a sort. Of a bird's eye view into, some. Of our own thoughts on addressing. Some of these challenges, and. So. I'll talk about a. Programming. Model that, goes. Across the continuum between, you. Know in some, sense enabling, applications similar to what, what. Moon was talking about in in, in being able to sort of share information, across, the. Edge as well as vertically. With a cloud and and. Then how you do do distributed, data replication and, then also I'll motivate some example, applications that. Might, use autonomous, H so that's, the. Thing that I want to do in the last in the in the next 10 minutes or so but, obviously, I cannot do full justice for, that and and, I invite, you to talk, to me or look, at some of these publications, that we have here so first let me talk about the geo. Distributed, programming model the idea is that you have. Sensors. That, might be generating, data and and, the application, itself will be a data flow graph that you can represent it as a data flow graph and and, what you want to do is you know Victor talked about what are the, sls right SLA, is in these kinds of applications tend to be latency, right so given the latency, properties, of the application, components, and the interaction, among them then you can decide how to place them in the. Both. Horizontally. As well as vertically in this in. This continuum, and and, so, that's part of what this, froglets, distributed. Programming model does is is. Respecting, the application, properties, it places, the application, components, in the in, the hierarchy, in the continuum, but, you know that's not a one-time thing because, you can have sensors movie and and, if sensors move then, the.
Faulkner That used to be processing, may not be the appropriate want, to continue the processing, which means that you have to migrate. The application, component, to a node. That, is close to where the production of data is so, that. Transparent. State migration is part of fog. Lisp programming model as well and and. Then. Multi-tenancy. Is something that I already. Mentioned. And so. Once you migrate. Something, to, a different fog node but you've left some, breadcrumbs. Over here and those, breadcrumbs, already brought over in order to continue the computations that's part of the programming model in terms of transparently. Worrying, about, moving. Both, the computation. As well as the state of this, distributed, computation, that's what this fog, list programming, model gives you and and. If you're interested in that please look at this, publication. And and, I'll be happy to talk to you also about that the second thing that I want to mention is the. Second project. Relates. To the. Recovery. Aspect of it so when you have data that is, very important, then. You want to make sure that the data doesn't go away when, there is failures. Right and and. So. So. One one possibility, latency. Is also a consideration, so what we want to do is we want to replicate the data in order to worry, about recovery, properties, so let's okay I go, ahead and replicate it in a particular way so that I can make sure that the latency, that is important, is preserved. But if I do this what's going to happen is if there is a coordinated. Korell correlated, failure that all of the readers lost right. And therefore, what I might want to do is I might want to replicate the data there, are far away from. The current location of where the plane is being generated so that I can have. Resiliency. And, recovery, properties, but that results in poor latency, right so there is this consistency, and latency. Trade-off. That exists and you know when. You think about some, of these frameworks, that are available in the cloud wanted, to let you do is it'll say well we're gonna give you eventually consistent, see that's all we going to give you and if, you care about strong consistency, you, look at it in the application if there is a violation of strong, consistency, you have to take corrective actions, in the application, and that to me is not very satisfying, so one of the things that my student came up with this idea of a, context, of interest so, a conflicts, of interest is that this, is a context, in which I want very, strong guarantees, in the data meaning.
That I want to know that anything, that I read is most up-to-date value. Right so so, that's, strong. Consistency, is providing, being provided in the context of interest but from the point of view of avoiding. Correlated. Failures what we're going to do is we're going to also replicate, it in a faraway no but if you don't give you get strong guarantees, on that so we give you eventual, consistency for, that so that's the idea behind. This. Fogg store work, and again only. A bird's eye view of the, details in that paper but if you're interested in that you. Can look at this this, paper so that's the second thing that is looking at the resiliency, properties and, the. Third thing is the, fact that you, have fog. Resources, and Mong alluded, to this that the resources are not you know multiple, racks you have limited amount of resources and and, you are having applications. That are continuously, generating, data 24/7. And and, and and, and and in that situation what you have to worry about is how to place this data that, is being generated so. That you can make sure that it, scales and it, also provides, you low latency, properties, all of you good, things that you need to do and so. That's the idea behind this data fog, and and. In particular, what happens is that there could be SKUs in the workload you know going back to the game example, that you had if there is a if. They're, soon, as the game ends everybody. Is leaving the stadium so there's going to result, a huge, amount of activity. Around that area and in that case the, edge infrastructure. Here is going to be oversubscribed, and so, what we have in data fog is ways by which we can say what, is the data that we need for, immediate query. Answering so, these are real-time queries can be answered by that and and, older. Data can be moved up in the hierarchy to. To. Ultimately to the cloud and so that's. The. Way we think about a, capacity. Conscious, data. Replication and. And. There, are details again on, how we do that taking, both the spatiotemporal, coordinates. Of the, location, of the data and then figuring out how, to sort of disperse the data in terms of replication, so details. Of that you can see in a in a publication that Kate recently came about as well so that's the those. Are the examples I have of things, that we're doing in the infrastructure, but I want to give you a couple of examples of. Applications. That might use autonomous. Edge, and, so. For this I'm going to give you an. Example of. Creating. Space-time. Trajectory, of. Vehicles. Right on the roadways now the, assumption, is that you don't you know already the, layout of cameras, on the roadways and and and. So. In principle, might seem like if I wanted to record the activity, of all the vehicles all the time it seems like a lot of data but but. In principle but but in practice what is happening is every, car starts, its, journey somewhere, ends it's somewhere right so it's an activity and the number of activities is bounded, by.
You, Know traffic conditions, and you know where you want to go and so on and so that's sort of the intuition behind how, we can do this even. Though we. Can do we want, to do this continuously, for. All the vehicles and so. The idea is that you, know if a particular roadway, you, know the, camera will attach to edge, notes, and, and. A particular, roadway, a site a particular vehicle and I can say through the Geo coordinate, of this vehicle and there's, a time at which I. I saw. This vehicle here and as soon as I know this I can communicate to. Other cameras. That, are downstream from, me they were given the direction of motion of the vehicle and the, layout of the, cameras, in the in, the roadways and I can tell the downstream. Cameras, that look out for this vehicle and I'm sending a signature of the vehicle to look for and analyze, the responsibility. Of the downstream nodes to three. Identify, that particular vehicle comes along and and, so. When that particular vehicle comes along it, can record that well I for this vehicle at. Geolocation. G2. And and. In. A particular time T V - and once, I have that once. This guy's confirms that yes I found it then. I can actually send this partial, trajectory. That I have because, this guy and get rid of it locally right, I don't have to hold it anymore and so. You can sort of aggregate. All of these trajectories. And. Keep, it together, and, and. Of course you can you can say well you know we, may not be able to keep, all the trajectories one point and you can do either. Lazy aggregation, and so on and talk about that in a minute and also you. Know as you keep on getting, all these trajectories. Accumulated. Here you might say that some of these old trajectories, and the recent, queries are for recent. Tracks. In which case I can even push. It into the cloud and and, not do. That and we can do grey lazy. Trajectory. Aggregation. So that you can keep the footprint, of. Storage. Requirement, in each of the nodes. Reasonable. And and. As. I said the, story is bounded by the activities, that is at a particular camera, and that's the reason that it is something, that is scalable and you can do it and and, again this, is only a bird's eye view of what. This. Paper does if you're interested, and talk to me and, and what we are doing right now is we're working with our. Campus. Security. Folks, you. Know I'm sure that this is something that most, of you experience, that every campus has cameras. And I asked my security. Guy what do you do with the camera we have a Google map and we can click on the camera and look at what is happening there's, no intelligence and and, we ask them what are the use cases that you're interested in and we can actually track vehicles. In, real time and then record that so that's sort of what. We're doing currently with that and another. Example that I wanted to give is is. In the context, of social sensing, and would. Be what, we have done is you know there you, have multiple components in the application, and again deployed, across as.
A, Connection. Tinh uation I can continue, and you can have network partitions, and if you have network partitions, then you can. Reconfigure. And resume, the operation and, and. And, opportunistically. We, can communicate when. It is possible using drones or buses, that carry edge nodes, and so on we can communicate with that and this is again something that. That. We've done, before and if, you're interested you can you can look at this publication, so so. With that I want to do and it's a bunch of ongoing work and I won't talk about that if you're interested I can do that because. I've been asked to wrap up and so. I'll wrap up, so. In the. Final comment that I want to make is that I hope, whether. It is propagated, or not but I hope I have communicated. That, we. Want to make the edge as a peer of the cloud and and. And and and Viktor articulated. How this. Is a disruption, that is coming. And and and, and. And in order to position, ourselves for the disruption this. Is think the thing that we. Want to be looking at thank. You. Kind. Of vision, the. Question I'd ask Sachi. Talks about Microsoft, as a company, that can be trusted with privacy, guarantees, I. Was. Anyway. I gave you privacy, guarantees, security, for data I. Was, sort, of expecting between the two tops so far that that that. Somebody would start to talk a little bit about the. Privacy implications, of, edge and of course here you are looking. At who might be driving a car where, they're driving to and who's with them that's, very clearly in this space so, I'm just curious to know what your thoughts are about. Very. Valid concern and so. For instance when, I talk to the campus, security folks they said well are. You going to publish this count with these camera feeds on on YouTube I said no I'm, just using it for analysis, and and, and, then recording, it and we're, not doing anything more with that but absolutely, right that you know privacy, considerations, are, super important, and we have a very strong group at Georgia Tech that is looking at security, and privacy issues I and I hope to partner with them in in terms of looking at some of those issues okay. On that note thank you very much again, I just one, comment I'll make to your comment, is that so, we've been dealing with traffic, cameras quite a bit and every, city is, sort of deals with it differently some, of them for example Seattle. This puts it on the web you can go and look at it anybody. Can, other places, like Bellevue will not put it on the web and then we have sort of signed some sort of agreement, with them so there. Is no clarity yet and how people, are dealing with it because there's no policy any. Way beyond, that real. Pleasure to introduce our, German Samuel, to you arch month is. No Street principal, program manager, in the Windows IOT. Azure. Group. He. Is in product, but, he has been in research he was part of MSR, in fact he and I worked together quite. A bit on some projects, you might remember like Hawaii. And then, a lab of things which was an IOT thing, but, then he moved on to the product division and now he thinks. A lot about IOT, and he thinks about the edge and in. Fact. Wouldn't courts owns the. Direction, in which it takes so, why don't him to share with you he has a PhD, from Purdue, so among his. Yes. I wanted to share with ask. Him to share with you our, existing. Current vision of where, we are. Thank. You Thank You Victor. It's a pleasure to be here Thank You Victor for the invitation, I've. Been working with a bunch of I, see very familiar faces quite, a few and so I've been working with the community for some time I. Moved, off to Asia IOT, which is the azure engineering, team, responsible. For IOT, about three, three. And a half years back. I, lead the azure. IOT, edge. Engineering. So. My. Team is responsible for, making, edge rail and the. One message I bring to. This group here is that, edge is really real there. Are actual customers, and I'm going to actually talk about some of those there are actual customers, who use edge, who, see the value in the edge and the cloud and they, marry, the two together so. There's a lot of momentum that we see out there and I'm going to talk, talk about that so when Victor asked me I should come here I said what. Better thing to do then, to bring in some real scenarios. Some, real actual, customer, use cases and talk, about some of the challenges, that they see our, customers and and, and then, really. Ask you the, community to solve some of those challenges. So. Let's make it real. I'm. Going to give. You a little bit of an, introduction into the IOT, patterns that we see, this.
Is How. We right. Now see of course the cloud is there and when, we make. A slide with a cloud the cloud takes. 75%. Of the space there's my design. But as, would also come here we. Look at IOT, s things, these. Are things around us they're things here there are things the, vehicle, is our thing smart, cities, manufacturing. Machines, robotic, arms they're all things things. Do two things they, generate data and, they. Can be actuated they can be controlled right, so these are things another. Thing so we look. At it in three ways his insights. When. These things generate, data there's a lot of insights, that you can get from their data you. Can understand, what is the state of the machine you can understand, the state of the environment, in which the machine is operating, you can also because. Now you have insights, you can have some actions, you. Can do some actions on the machine on the thing or you can take actions, on your end you can adapt some processes, and so on so, that's how we look at this data, of course flows as we, see it today too, from the things to insights. In the cloud and from, there to actions, and then maybe actions, come back to the club due to the thing but. Then now if you put an edge. Concept. On top of it you. Have I'll, skip, this. You. Have the same thing, you have the same insights, you have the actions and you of course have a cloud gateway but what you now do here. Is you. Move some. Of the insights, you get from data some, of the actions, that you need to do and you bring, it right there on the thing now. Your think the robotic, on the huge. Milling. Machine your, car your, smart city your your environment, like this conference room they have become smart in themselves. Because those insights, which essentially, is AI machine, learning parts, of those have been moved to the, thing and that's, how we look at the edge because. Edge and, the, cloud work together now. I won't go over some, of the virtues of the cloud and some, of the virtues of the edge only. To say that both. Of them are important, in their own self. You, cannot, do a huge. Machine. Learning, training exercise, with tons of data from distributed. Nodes coming, and you, only need the cloud to do that because of infinite, near infinite resources, but, then at the same time if you are looking for near.
Real-time. Latency. You are looking for control, loops which can run reliably, right, next to the machine you are not going to send calls up to the cloud and. For them to come back this. Might not be reliable you need those insights. Those AI, machine, learning right on that machine now. We. Can talk a lot, about these these virtues of the cloud and the edge and how they sit together and how they augment each other but, the way we have done now coming to what we have done the product is to, create a consistency. Between. The cloud and the edge and by, consistency, we mean the consistency. In programming, models, consistency. In the application model. What. Does that mean that means that if there is a workload that you are running in the cloud. Imagine. A stream analytics job, which is ingesting, tons of data analyzing. The data and then coming out with some statistics, the same application. Model, we have created, for the edge which. Means which has which, has implications, beyond. Just the model which. Means you can actually move that same workload without. Any change, from. The cloud onto. The edge without. Any modification, without writing any code, all through, configurations. So, that is the consistency. That is how consistency. Between the cloud programming. Model and the application, model and the, edge programming. Model and application, model helps us it. Has implications beyond, just technology it, has organizational. Implications, right. So now if, you, do not have this consistency, in the programming model you, need, developers. You need people who actually program, the cloud, application. And then you need, people who actually go in and program the edge, application. But, with consistency, and an infrastructure. Which allows you to move, workloads, from the cloud to the edge you do not need that you just need cloud developers, who develop, their their, applications, in the cloud make. Parts of it which they want or they think or they optimize, to run on the edge and simply, configure it that way and then they can move it back so that's. Really how we, look at the edge that is what azure. IOT edge does by, the way this is azure. IOT edge there's, also another, product, the, edge browser this is not it this, is edger IOT, edge which we are talking about. Okay. So now I'm going to make it a little bit real then I'm going to stop stop. A little bit and go into some design principles here. This. Is a real customer that we work with they have a mining operation they, have spread these, pumps, which are spread around, this. Very. Remote locations, many of these remote locations are, in the ocean they're, not even on land and then, they have one central, place where they're controlling them network. Connectivity to these remote locations is very bad or very, expensive, so they typically would use satellite, connectivity they. Can only get alerts otherwise, it becomes very expensive so.
What, They do and also sorry, and also. Reliability. Of operation. Of this, pump. Is extremely. Important because, if it goes down you lose money, you lose oil which is mine so. This. Is their, world there. Is a pump here which, is going up and down similar. To that one the, pump has really. It's boiled, down to two values, of pressure pressure. Above. The plunger and pressure, below the plunger and then. This on the x-axis is the position of the pump, so, at all times there, is an envelope of operation, for these two pressures the. Minute it goes out, of scope, out of that envelope this pump fails and this, whole, operation goes down now. Obviously there is a requirement for it to be near-real-time, so they need to monitor this there's also a requirement to be predictive. You. Cannot wait for this envelope, to be broken to. Take some action you, need to predict, when, this would happen so that it does not break down and you can do something about this power, and. You want to do it in offline this pump could, be operating in the middle of the ocean and you do not have any connectivity, to the cloud now. How they do it is right, now they. Would have the pump they actually have some, automation. They call it there's a gateway there's, a SCADA controller, there's an HMI right next to it and then, when, that and then there's an alert there's a supervisor sitting, in the central place there. Is an alert which is generated, when that and will is broken and then, the supervisor. Would send somebody who would start tweaking the pump shutting, it down or doing, things which they can do to mitigate that that i
2018-08-31