 Good morning to all of you and we see that this is the time to start 9 o'clock is the time so even though we expect lot more people to come but we will start initiate. So the topic at our hand is driving cloudlet into OpenStack. So the main at the end of the session if you want to carry some message you remember cloudlet, cloudlet and cloudlet that's the message you need to carry with you. Given set that since we are here three of us are here I have Professor Satya Mahadevan, Mahadev Satya Narayanan and then we have Ralph Schuster these two are the fellow one of them is the I call visionary other is the missionary I am the executive on their behalf. So that's what we see and all three of us have Intel insiders because these are the four people who are the vendors like Huawei I am from Huawei he's from CMU he's from the service provider that is the Vodafone and all of us have Intel inside and that's the four people are collaborating parties who have come together to drive the cloudlet. So I will let just give you introduce a brief introduction not much just to see I have been this is my seventh conference here some it probably I am the lucky seven so I am here on the floor otherwise I will most of the time I was on the other side and I am in located in Santa Clara and I work for Future Why that is which is Huawei R&D and my I have been working globally starting from India to US to China and I am here in front of you hopefully we will do something good for NFE and SDN which is our vision for the open edge computing. Dr. Ralph Schuster he is the head of innovation center Vodafone he deals with the research he provides the incentive for us to develop newer technologies and as a service provider they have been very helpful to Professor Satya also and he has been CTO at startups and all and very accomplished person from Europe and I think we have been blessed to have him in our team to lead us for the actual mission of cloudlet. Then this I do not need introduction process there he has been in the Carnegie Mellon University he was the one who wrote the case of VM based cloudlet for mobile computing that was the one 2009 when he started on this and today we see the reality of the cloudlet at the edge. He was the original founder of the Andrew file system so he has been in the distributed file system distributed computing and several of his students have been all over places from Robbox to everywhere so he's a eminent personality and we like to see him bring to us the vision of cloudlet and this is the time for me to hand over to him. Thanks. I've got this thank you very much Prakash so what I'd like to do is to just introduce you to cloudlets tell you what they are why they're important why you should care about them and then I'm going to turn it over to the rest of my team to take it forward. So what is a cloudlet? A cloudlet is a small data center think of a enormous data center that is shrunk into maybe a suitcase sized maybe a rack maybe even smaller but the critical properties are that it's located one wireless hop away from mobile devices so it's very very close the latency is low and any wireless improvements whether in latency or bandwidth immediately translate into application perceived improvement. On the other hand a cloudlet does not suffer from some of the challenges that mobile devices have to suffer from you don't have to worry about the weight you don't have to worry about the battery life you don't have to worry about small size heat dissipation all of these challenges are challenges you have to cope with when designing a smartphone a Google Glass or any of these other kinds of system so as you'll see in the rest of the stock it's a catalyst for brand new mobile applications. Here's the deployment model that we envision for cloudlets at the top in the cloud are unmodified cloud services what exists today and at the edges located very close to mobile devices are what we think of as internet infrastructure in other words you don't have an iOS cloudlet you don't have a Android cloudlet you don't have a Windows cloudlet you think of it the same way you think of Wi-Fi infrastructure it's a cloudlet and for the same kinds of reasons of multi-tenancy isolation etc that matter in the cloud those same factors also apply at the edge so we can see coexisting possibly untrusted back ends of different mobile services running on the same cloudlet so one way to think of what this is this picture is is you're all familiar with CDNs Akamai made them famous but many companies have them think of cloudlets as defining a CDN for computation whereas CDNs originally were created to do data caching cloudlets do for computation what CDNs have done for data so why should you care about cloudlets in one slide let me try to summarize for you there's a lot of detail on this that you can look up and we have plenty of resources to share with you but let me just distill for you the essence of why cloudlets are important first most important reason is latency I'm going to drill down into this in just a second but it's also important from the point of view of bandwidth imagine all of you in this room wearing GoPro or Google glass or some such device how much would the total cumulative bandwidth ingress bandwidth of all that video be if a significant fraction of Tokyo or any other large city and whether it is mobile or fixed infrastructure the total transmission of video data into the cloud would be very substantial what cloudlets let you do is to push the analytics to the edge so the video only has to travel as far as the cloudlet is analyzed there and it can be stored there in raw but encrypted form later even examination of the extracted analytics suggest that you should go look at the data you can do that at least for a certain period of time that you're able to buffer the raw video so bandwidth is a second critical critical reason the third reason has to do with privacy it is the case today that concerns about privacy are starting to be a big source of slowdown in the deployment of the Internet of Things if your water consumption can be monitored and through machine learning begin in for every time we flushed you flushed a commode in your home I can therefore infer all kinds of patterns about the state of your health it's not something most people want to reveal to the world but that's what's possible today the ability to place privacy enforcement software close to the point of data capture is an important capability that cloudlets provide and last but not least cloud computing presumes the existence of connectivity to the cloud but as you all know with the emergence of denial of service attacks in various other forms of attacks the ability to get to the cloud may be compromised in the future what you need is for cloud services something like what you use for power you purchase a uninterruptible power supply you plug it in and for brief periods of failure you can continue operating cloudlets provide you a platform on which those cloud services that have been suitably modified can use the cloudlet as a fallback location so these four critical usage models of cloudlets are going to drive its future but latency was the most important and original motivating reason and that's what I'm going to focus on so does it really matter let me give you the answer suppose you're writing an augmented reality application use your smartphone use the sensors like the camera on it or the accelerometer you transmit this data to the back end software and then it is analyzed there and some result is shipped back for example the name of the building that's in the image is an example of a simple augmented reality application you could run the back end on a cloudlet connected by Wi-Fi or for GLTE or you can place it in Amazon and Amazon data center either in the east or west or Europe or Asia or you can run the whole thing locally all of these are possibilities now these experiments were conducted in Pittsburgh and the closest Amazon site six milliseconds away is Amazon East Amazon West which is in Portland is about 80 milliseconds away and the others are much further so look at this curve what it shows here is in milliseconds on the x-axis and here is a fraction of the results that come back sooner than that many milliseconds so if you're looking at 80% of the results using a cloudlet come back in less than 80 milliseconds whereas if you use Amazon East 80% of the results come back in about 250 milliseconds and things get progressively worse as you can see and running it purely on the mobile device even though there's no wireless transmission involved is still worse the reason has to do with the computational power of the smartphone relative to what you can put in a rack as a cloudlet the picture looks roughly the same but shifted to the right if you use for GLTE but in all other respects the results are quite consistent now this is what the user experiences in terms of crisp computation basically if you use the cloudlet the augmented reality app is really crisp and snappy and that's the user experience that you want however using a cloudlet is also enormously beneficial in terms of battery life in terms of the amount of energy consumed on the mobile device to perform the operation so if you look at this if you look at augmented reality which is the other column performing the entire operation on the mobile device for one image eats up 5.4 joules if I ship the image to the cloudlet I consume Wi-Fi energy transmission costs and reception but I don't pay the cost of computation which is done by the cloudlet the total amount of energy used is only 0.6 joules dramatically lower if I do the same computation and Amazon East identical first wireless hop just the fact that the result takes longer to come back causes the energy consumption to increase by a factor of 5 so the savings are huge of course it depends on the kind of application and you'll have to do the experimental analysis for your specific application to understand but the big picture is pretty clear so so far what I've described to you other than augmented reality are pretty well-known well understood application classes what new applications are enabled by this world I want to give you just an insight into one new class of applications so imagine you are wearing some device mobile device like Google Glass or HoloLens or any of these others augmented reality devices everything seen by your sensors your camera your microphone your accelerometer is transmitted to the cloudlet it sees what you see it hears what you hear but faster than you can think it computes on that sensor data and whispers in your ear some useful guidance that could possibly help you think of GPS navigation maybe to get you to this building use your GPS on your smartphone or in your automobile think of the step-by-step guidance that it gives you could we expand that to other kinds of tasks and other kinds of sensor data that is the world that this opens up let me just show you one example to impress upon you the significance of latency of how this could work there are many example applications we've built of this kind this is just one our Google Glass based ping pong assistant helps the user to play better ping pong by whispering whether the user should hit to the left or to the right the hint is based on the observed ball position and opponent position okay you get the picture how did that work every video frame is analyzed using computer vision to find the ball and the opponent and then it is compared with the position of the ball and the opponent in the previous frame from it is calculated the trajectory of the ball and the trajectory of the opponent and based on that in real time you get guidance to hit the ball to the left or right simple application but I just want you to think about how much computation is involved and how critical real-time responses so this is the world that high intensity compute close by can do for you this is only one of many future use cases we envision think of going to IKEA buying a kit as we're getting printed instructions giving step-by-step guidance troubleshooting industrial machinery is often a very complicated task trying to get guidance step-by-step would be very powerful medical training you know that's not even a real patient there that's just a mannequin to allow a student who's this guy to learn without this expensive doctor standing by is a kind of learning that is possible with this for elder care I know that in Japan it's a significant certainly in the United States it's significant the use of medical devices to get your blood pressure sugar level etc etc is very important keeping healthcare costs down but if you don't install these and do the readings well the readings are meaningless and so getting guidance to do that is important and then my favorite is to help you lead a healthier lifestyle just as you're about to eat that donut to yell in your ear stop don't do that okay so our vision is that open-stack on plaudlets whether it's at home in your car in the cell tower in an aircraft it's the way that we're going to achieve the vision described and to tell you how I'm going to turn it over to Ralph so good morning to all of you warm welcome first of all I would like to thank you for having me here considering that I come from a different universe from the telecom world that I'm a complete newbie to the open-stack summit right so thanks for allowing me to talk let me expand that a little bit this vision we have just heard a little bit also from a perspective from the telecoms industry right what we would like to enable is really put cloudless everywhere as Satya just said we would like to put it in base stations we would like to put it in core networks DSL boxes Wi-Fi hotspots and basically offer storage and compute power everywhere on the planet right next to the user I think this is the key that we believe that low latency and the other drivers that have been mentioned make the need and form the need to actually put that edge infrastructure in place and to be clear this concept has to be open right it has to be open also in the sense of that any application can come and use it it has to be open in the sense of that is bearer independent whether it's Wi-Fi whether it's 3G 4G Zigbee whatever this infrastructure this computing storage at the edge has to work for all of these situations and this also explains why we believe that open-stack is the key component in all of this because what we would like to offer is a platform that is a multi-tenant environment that any application can come and can utilize and therefore we believe that open-stack is the ideal candidate to actually play in that in that environment now we have talked already about applications you know we we can even further extend the list we have just heard I just wanted to mention maybe two other areas that have not been addressed that is for example assisted driving imagine you drive a car and you get information about the road in right in front of you in real-time from the roadside and you maybe even get a an overlay on your windscreen allowing you to look to look through fog to look through trucks driving in front of you or just get warnings at an early time we can also imagine drone control you see all of these little things flying around everywhere but you also need information about the direct environment in near real-time so last but not least also I guess online gamers get more and more sensitive about latency we can imagine that they they set up their local edge server and their local cloudlet component in order to get the best performance now from a business perspective right who are the players that that should act here first of all obviously there is a communication service provider like waterfall in the company I work for you know they provide the infrastructure and the networks and all that above that obviously there has to be an edge cloud service provider or that offers components at the edge and manage it from the back end we will need some central components like a registry of cloudlets being available and their capabilities and also obviously the application provider has to come into play and has to offer a 3d architecture here where he has an application component sitting on a mobile device or whatever this where this application runs on where he has a component running at the edge component and then the back-end cloud components right and now the question is you know is their interest is their value in edge computing for all of these players and if we have done the value chain analysis and we come to the conclusion that actually all of these companies and all of these value chain segments have an advantage for it if you look at the telecom infrastructure providers and the the cloud technology providers I mean they have definitely an advantage in terms of you know new products new services innovation image and so on network providers like motor phone like my company they are very interested because they they can extend their offerings into the cloud space I mean they know the they own the edge and the network anyhow right why not putting cloud infrastructure into that and monetize it and then of course the application service providers and the developers application developers themselves they can offer a superior and actually disruptive changing customer experience based on edge computing so for overall we feel that this proposition has advantages for for all players in the value chain we just have to bring them all together and make it happen now in order to make this happen we have formed an open-edge computing initiative really with parties from the telecoms world service provider infrastructure providers and we also reach out to IT players now to really to really drive this ecosystem around edge computing because it will not happen by itself or there's a bit of a chicken egg problem if low latency applications on there infrastructure will not be put in place if the infrastructure is not in place the application developers don't work right so we have to bring all the parties together namely the application providers infrastructure providers as well as the service providers and you know make sure that this is a coordinated development we drive and as part of the key targets is really of this initiative also reach out into into your community into open stack to say you know don't you want to join that initiative in terms of you know bring open stack from maybe a data center environment basically everywhere on the planet and you want to enable open stack to run also in environments maybe that that has not been thought of in the past so this is the the invitation to all of you to say you know is that a sensible target we don't know right we want to listen to you it's a sensible target to to you know utilize open stack also outside of data centers we will also reach out to application providers this is very important because if there is no applications you know there is no business at the end of the day we have started already developing demonstrators we have seen this ping-pong assistant and we strongly believe that this is the right way to to engage with the application provider industry and explain to them the the opportunities they could have here obviously we will be as with operators and with cloud service providers in order to bring them to the table as well so this is the open edge commuting initiative now currently we have Intel Huawei and Vodafone as key players in their Carnegie Mellon is the academic partner we have and but this initiative is open to other companies to other partners you know those who really have to come willing to compute to contribute significantly to it but to be clear this is not just these three companies behind that there is a whole initiative in the telecoms world already there is I mean this is how telecom industry works if they see a new opportunity they start standardizing right so there is an Etsy standardization group called mobile edge computing 50 companies are currently playing there mainly telecoms companies but also some IT companies are there you know and they very actively drive that forward they see this also as as one step into 5G the next generation mobile networks where the latency is coming from 4G latency of 10 to 15 milliseconds on the air interface down to 1 millisecond so basically they see this also as a step in in that direction and just to say you know this is not three companies trying to do something this is a whole collection of global companies that would like to drive that but what the open edge initiative tries to do is really to say we reach out to you guys and to the application industry to make you aware to bring you into all of this and to make you involved now just quickly my last slide before I hand over to the technical part we seek your support here right we believe OpenStack is the platform in the cloud world and we would like to make it platform of choice for edge computing and hence you know we are here we we offer that to you you know we have developed something and you will hear that details about that in a minute but we seek developer support you know people who who help us making this really integral part of OpenStack that any any OpenStack release is enabled for the edge is enabled for applications to actually discover it and actually utilize it we will as the next step we want to develop a reference platform that actually does it offer the services tend to end and at the end of the day we have to ensure that there is a smooth customer we also invite companies to join right and to help and just as a as a as a main message to you you know join us to bring OpenStack from the data center out into the world basically everywhere with that I hand over to Prakash who will guide you into the implementation we have already done thanks thank you doctor of and professor yeah now the vision portion and the mission portion is over let us get into execution so what do we have here is how do we execute so the first thing we think in this world of OpenStack is API API is API is what do I have so we do have cloudlet and cloudlet which is visioned by professor Satya falls under open edge computing currently that's why we call we have to name an API so we call it as open edge computing API which consists of the open edge computing cloudlet API which is related to the computational and storage aspects we believe and the communication API which is related to network because we say network is computing be that so we need to have that networking to allow different APIs to interact with each other so given that we start with cloudlet API so we have to think of that back end being Himalaya and this being the Fuji mountain so something like a huge one and a closer one so whatever is closest to you you connect to it from the device so you have a device and you want to connect to the edge rather than connecting to the back end which is far off so if you connect then once you connect then you have to start engaging in the API is what do you do oh you want to create a VM well why are you creating know what does it yes know what does it but know what does it for is standard of the shelf and we do want to standard of the shelf but we need little more than what it does in terms of latency bandwidth in terms of privacy in terms of security in terms of the various values we professed so that's the reason that API should not only be just simple API of like if you have a cloud and you can create everything in the cloud using Noah I can create a VM I can take an application push it into it I can run it then why do I need this not really because we do have certain constraints which we profess and we want to attain that therefore this API which is the cloud-led API we see cloud-led API we mentioned those have to contain those elements and that's the that's the element which we are pointing here and the so the other portion is of course to the com layer the communication API so we are talking about the device API because there is a client portion of it without a device what do you see a user having a device can access so therefore to connect to those applications which are very constrained we use this device and obviously as a service provider if it is everything is device managed I trigger at the device and then get everything done and then I'm done then easy but the other end somebody has to pay for it somebody has subscribed for it somebody pays for it so there is a management aspect and that's why the service provider would like to manage from the back end either in the cloud as a registry or at the edge with a registry so that somebody can charge for it measure it so those are the aspects which the API should cover so moving on I will now get into what APIs we do have currently so first thing is this working piece sure it's a working piece we as CMU started this almost in ice house Juno and now they have pushed into kilo and our goal is it's you know open stack moves six months it is so fast rate that by the time you blink it six months is over and you are already falling behind so our goal is right now chasing for last three open stack releases we have been chasing and we have come to kilo now liberty is released on 18th and so we want to push to liberty but before we move that what do we have in API terms so the bottom two API are the most important APIs VM synthesis and VM handoff these are the two which are live these are offline other three are there those are supporting APIs but the bottom two are the main so what is synthesis and what is handoff so if you look at a glance cat catalog you have image so you have a image so we have a base image of course we don't want a very very big image we want a image which is footprint-wise is small so that we can easily use that that's one criteria the other criteria is once you have that how do you store it okay and once you have stored it if you if you run a VM in no one terms if you run a VM that is instantiation once instantiate you have to run an application on it so what is the image if I take a image image in terms of computer science is either disk with a hash table and then similarly you have got a memory with a hash table now that's the process view of the VM as such as a image level now I run something on it it creates some additional context so those context could be compute context storage context network context these have to be preserved in that so the delta is the overlay that is if you run a standard gold image and then on top of it you run a process that process creates something delta that is the delta we need to collect so that's what so normally in data center what we do we do a live migration here is the VM migrated there it starts running there of course if it is live well and good if it's not then you the user's perception is they have lost something so that you should not lose something that's the live migration but in this case what we have done is there is a format in which the VM is since so you create an overlay and that overlay has four components disk incremental and memory incremental and then when you run on top some application and then synthesize it back then that is what is the recurring the snapshot so you have a snapshot you recover you add some process and that delta is added and then then you want to move it like I am moving from here to there my cloudlet should move from one edge to another edge for that to happen I need to run it stop it take the delta push the delta to the other end started new VM and then add the delta and by moving your changing locations so you are going to lose some of the what you call the state network states and all and so we have to make sure those are all preserved and this should not be pursued by the end user from the device that there is some effect of that like if I am not able to see something I am losing something then I have not made my latency bandwidth sufficient to make that happen so that is the goal behind this so I will just move on to the next that shows actually that this is what we had in kilo and this is what we propose that we want to have like we have not completed registration we want to make sure it is registered and then we want to do some additional stuff like create a edge group so that we can have two VMs while moving if it is a group if one fails active passive passive active those kinds of things are included and obviously we want to attach detach or we call it as link and link so those are the certain APIs which would like to have okay and that's the goal and let's see how we move on that so the future is much more than what we have described but at this stage we have to be grounded that the basic what we have is what we have and we would like to see that this also goes into the the APIs are able to be exploited in the future in the 4g and 4 and half g and the 5g which is coming and that will require some kind of a state management and that's we will just I'll skip it this is but this is required as part of it this is what we have state of OEC current so open-edge computing cloudlet currently in kilo adopts the mobile edge computing concept which is there in the HC and the OpenStack APIs are there it is running through the NOAA API currently and it is using the messaging RabbitMQ to send the message to the compute and in the compute you have the drivers being replayed NOAA driver is being replaced by cloudlet and then it is also managed with through the hypervisor in this case we have used KVM so whatever Libvert has got it has been what you call the class of service it gives that has been inherited by the cloudlet and additional capability like synthesis and handoff are added to that so this is what the API is and this is how currently we are managing and for the for the sake of time I would say that at this time I would say that we are at the end of it but we do have Kiryanka can you raise your hand Kiryanka is the architect he is in at CMU and he is doing his PhD and he has been the architect of this he has helped us build this and we would like that there are some challenges he faced as you can see portability of VM is one of them which was handled by him like you have different Intel processor if it is sandy bridge or it is something else then you have that if the hardware changes at the bottom then the states are not maintained so we have a portability of VM issue we similarly have a network stage issue whether you can state can be maintained or not maintained based on the MAC address changing or the IP address changing so those are the challenges plus we also need quemo and KVM which are not up to date for us to really use it efficiently for low latency that needs enhancement and the testing of multi nodes so when we handoff we have one open stack one VM another open stack the same VM to be recreated so we need credentials etc etc so those are the challenges we are facing and we look forward that we should be able to sorry we should be able to get this over and our plan the last kilo of time frame was we tried to do whatever best we have within our own resources and we look forward to this DevOps support from the open stack community and the essence of this is this is our goal we want to deliver open stack in next Mitaka to cover the gap from Liberty to Mitaka and for that we need your support so anybody who is willing to participate we went to the open stack what you call the compute team and they told us that we should better have our own cloudlet as a new module but to do that we need lot of preparation lot of help from the community and we look forward to all of you who are in the attendance as well as other companies who would help us in doing that and especially the application service providers as well as the telecom providers and infrastructure for all of them the welcome to help us and I think at this stage I would just give time for the Q&A and here is the detail please ask any questions and we have yeah go ahead Susan so yes what it means is we do have NOAA as a project we wanted to tag on to NOAA as a microservice API so that we can just use existing API however they consider that changes to the API will impact the development in the Mitaka cycle what they are in they suggest that why don't you send a start a new what you call a module called cloudlet and we are consulting Kaili mystery and other people and we hope you'll be able to do that question sir can you can you go to the mic please everybody go to the mic for questions please line up if you have yeah hi Nigel cook the question is the API's that you have all the lifecycle operations around the amps and so forth how does that relate to the interfaces that are exposed in NFE I so they're also you know have a sort of a similar bent because of the telecom flavor is there any overlap there or are these not covered in that API set so the question is the NFE API is whatever they have interfaces if I under if I interfaces and then those are those reflected in our implementation so the answer to that is open stack de facto standard doesn't follow the standard what the HC community defines however op NFE is the platform which is trying to bridge the gap and as a open edge computing we are still very early in the day we just want simple API so at this time we are not trying to address any gaps we hope op NFE will be able to cover that for us and then we can probably leverage on that so we are at a very nascent stage at this stage to consider but we do follow that we are looking at mech to provide the normative API references and they don't have so what we will do is we will provide and try to influence mech to adopt that that's what we see right now does that answer your question it does thank you okay any questions from the yeah go ahead yeah go ahead no come come in front instead of packaging the apps in VMs have you considered containers are very good question okay put it this way we do have roadmap for that but at this stage in this year in 2015 we are not taking him but 2016 maybe March time frame is what we plan to do that we do have a roadmap for it yes you are welcome to join us in helping us yeah no we seek participation from the duty we nothing runs without containers these days right and the faster the better yeah so the security aspect is something which we need to address but otherwise I think we are fine with that any other questions please are we running out of time are we in time we we are in time done okay so I think if in the absence of any questions I will just announce few things before the first thing is there is a HTTP colon open edge computing dot org Ashley Ashley at the back stand up Ashley Ashley said so Ashley has already it's inaugurated this you are most welcome to go to HTTP open edge computing dot org we have all the details for you including the references to the codes and what we want to do etc and we thank you go ahead you have anything exactly exactly there is a KVM for NFE project under op NFE which deals with the portability of VM and etc in computing sense and that's led by Intel and other folks and we welcome everybody to join that to help get there and then there is another thing which I want to there are many people behind this don't think that we are the we are only the showpieces the actual work behind Ashley has managed to do the open computing edge org site then you have Kiryanka our architect he will be available he will demo if you want to see the demonstration you will be available in Huawei booth P14 at the marketplace at 11 o'clock you can watch the thing what the code is what is our state of under kilo what we have developed and how the POC is done he can show you the migration and all it's a cool thing to look at and if you want to join the coding and all Gunther class is not here he's in Vodafone but he is also one of the architect behind it and he works with Rolf in the Europe and then we have Patmanism Pille Babu we call him Babu is a great supporter of this and he is on the compute side he has tried to see how to accelerate how to get the latency lower and lower and lower that's the target of 3 to 30 millisecond we say 30 millisecond for LTE 3 millisecond for the Wi-Fi we want to push it sub millisecond if you can that will be the 5g then so we have great support from all of the players here and we look forward to your participation and come and join Kiryanka at 11 p.m. at the P14 11 11 a.m. sorry 11 a.m. I'm still there okay 11 a.m. P14 Huawei booth marketplace please thank you very much I think we should quit the other team is coming in and thanks the open stack community for allowing us to do this sorry for not mentioning