 Hello guys, hello everyone. What an amazing day we have. It's my turn to do the final presentation Basically, let me just sum this Okay Yep Again, what a wonderful opportunity we had today a lot of amazing presentations. I'll try to keep it up And for that I brought today's topic which is migrating 5d core to the edge Wisely and confidently with Kubernetes Many of you have heard of 5g already Kind of sick of what if 5g maybe it's just some A trend we are going through these years or the field pastime But believe me and it's the fact that 5g and Kubernetes of they have a long story And I will try to show into this presentation The contents for today. I want to try to be quite a Simple of many of the 5d topics One introduction then I'm going to mention the problem the problem for this presentation. What is 5g edge? What is mech if you have heard of it? What is the strategies that Kubernetes? Can that can be applied in Kubernetes to achieve this 5d core at the edge In both at CD performance put scheduling and At the final I'm going to do a demo analysis. It's not going to be a full demo I think you guys all all of us wanted to just wrap it up. So let's try to do some summary A little bit about me my name is Marco on cell is I'm from Lima, Peru. I'm based in Japan since two years three years ago Working as a 5g solution architect at the next in Japan Oh, basically Doing all about 5g rollouts in Japan for different telecom operators big coffee fan and Apart from my current role. I also am part of the AWS community builder association. Basically, we try to Talk about what is a double a solutions in my case I try to make blocks about networking content delivery services and how to integrate cloud services in general One half time, of course, I tried to do some hiking I tried to take advantage of the amazing landscape of having Japan So for the problem statement, I want to just make a simple questions Is any of you guys are working with 5g or 4g currently Kubernetes or have ever worked with 5g or 4g? Oh Right to So I'm going to do something really simple I'm not going to dig in deep into 5g for for us for the presentation is going to be just some Obligation that is going to be run at the edge so basically edge I can define as a Scenario where we have a data producer trying to send a bunch of data and in order to get a response we are Trying to put the server close to that to the data producer. That's the main concept of edge solutions of course, we have Different scenarios computing of loading Data caching and that are processing and so on so forth But when we try to do this there are common of important question we we have Is this solutions is actually going to be profitable because at the end of the day 5g is one of the Solutions for enterprises Therefore the question is massive IoT. Is it really going to help me these 5g edge? It is definitely going to handle critical workloads. Is it going to have some redundancy? so that those two questions are appears many forums when when we talk about 5g and the edge and the most important Is it saving cost? Is it being profitable? I would save me money with this I would make it. I will be spending more money and Faster TTM or time to market 5g core and 5d solutions is meant to be meant to have a fast delivery to the customer because these days obviously customer wants everything Yes of the speed of light. So that's what I expected from 5g I'm going to quickly talk about Timeline so this is gone. This was crazy 5g and 5d edge has started in 2019 and in just less than three years guys. We have been we have a bunch of rollouts commercial networks and amazing R&Ds projects all over the world Germany USA China Until the last year is more than 660 million 5d subscribers and definitely when you see that number you can think about it This there is some use case that we can we can have with the 5g particularly in Japan. I have a cool Research that there are indeed some R&D R&D projects especially with well-known telecom operators and the main business case So we have is machine-to-machine, which means a device is sending traffic to another device is having a machine-to-machine communication peer-to-peer As I said before 5g edge it's It's impulse is promoted by the fact that there is a huge investment worldwide more than seven billion dollars of just Forecast for 2027 and of course we are talking about edge intelligence overall so let's Deep in this big into what's 5g edge and make or multi-axis edge computing is For a sake of the presentation don't care about these blue boxes. There are a lot of information this talk about infrastructure, let's talk about what makes 5g at the edge works and If you want to focus on that, let's only talk about the red boxes red color boxes This is a one a standard Architecture basically it's defining how different operators trying to apply these use cases application computing of loading multi-user multi-network application B2x which is vehicle to everything and Auto-mobile auto-mobile workloads and so on so forth many use cases. So let's dig into those two red boxes Okay, Kubernetes and 5g edge Kubernetes 101 you guys know what it is, right? It's it's one of the key topics of more basic topics We have a truly consistent distributed keep all your store and use the rough consensus protocol I was surprised that my colleague from I call it the speaker from Hitachi talk about this Consistency protocol and what are the challenges? Indeed. There are some challenges. We are trying to solve with 5g core There are different use cases in the presentation that talk about financial bank solution, right? It's true consistency the data must be getting both ends That applies for a bank solution if we move on to the 5g edge IoT which mean devices there is no such Priority or have this strong consistency, but yet having this a strong consistency makes a Huge sacrifice for the HCD Let's go to the this small diagram Every time there is a pot scaling out every time I'm trying to the cluster is trying to collect a new pot There's a lot of bands of Communication between the AP server because the controller and in between each of these steps We have the famous data store at CD What happened? Every time we do one call every time we do one operation get or to update There has to be these famous strongly consistency synchronization if we have three that the store as you can see in a presentation there's going to be Synchronization between these three nodes before Replying to AP API server and this go all the way until we schedule a pot with the least about if we have Container registry we pull that the image and every time we have this Strong consistency Operation you may think but Marco this is just common Kubernetes. Why does affect us? I mean, it's just a few milliseconds few seconds overall. That's true But let's think of this way Apart from 5g at the edge. We're also trying to use bare metal We were trying to not relying on one vendor which we're trying to open the different options with a double S I should we would cloud today I saw a bunch of presentation just talking about those public providers and You guys know that they charge for everything traffic, right every in our traffic they charge So every time you do this synchronization you are pretty much increasing your monthly fee So that's the one thing that is not mandatory for this use case IOT What are the strategy in this case? Well, this is kind of the opposite of what my The previous speaker mentioned. It's move from a strong consistency to a lazy thinking for that We use the conflict free to look at data types, which it's not as As that name says it doesn't need to perform immediate communication with our notes instead, it definitely Answers the query or replies the API server and performs this all these two Replication state-based or operation-based of course the first one is refined like a image copy between the notes and the second one is just Cloning the commands to the different data stores obviously the second one is more bound with efficient So it's one of the strategies we can use in order to reduce this Delay and improve the overall cluster performance Of course, I'm focusing this solution on the massive IOT scenario Is it one of the keys now that we have so far worldwide in 5G gore edge Second strategy again basic concept how to do a post strategy guys how we Excale we do the basic HPA via BPA and a cluster of scaling those three options We have been working. We're seeing all our different labs different Sessions, so it's very common But if you try to take this just as it is to the 5G core you have one challenge 5G users are not that predictable for instance You guys know like 5G It's meant to be to use a many use cases Video gaming streaming Tiktokers biggest tool. So one user can do many things So it's not like a one user one traffic. It's not one user one request. It's a one user many requests So if you think that this is going to solve your problem, you may end up having an outage So that's what that's why the telecom operators don't want and for that one of the strategy it's Develop and improve the proactive force of the scalar. This is obviously with a research If you can see the diagram on my left There is a kind of like a more aligned to what is the Machine learning Intrusion to the 5G for instance We have a one control loop in this case for the sake of the presentations every 20 seconds is going to Evaluate it's going to receive metrics, but not only CPU memory storage Not only the common ones more data that is aligned to the clients. It's aligned to the use case That's a big difference between 5G and let's say any database server. It's not that straightforward So based on that you can predict the behavior and the take apart action. It's not just I'm going to increase when I found the issue that In a telecom service is just service impact and is customer complaining its impact to the to the brand So that's something at least so far is one of the key discussions for the 5g edge One of the benefit of course, it's a proactive workflow for out the scaling Multimetrics that can be used Text into consideration of course limitation and constraints of the resources. This is important because As I mentioned before 5g edge is not going to use one simple Brand or one simple flavor. It's going to use multi edge for cloud providers It's going to use as I think today was presented bare metal so different options So for this randomness for these variety of options We can use this proactive both out skill and we can use the benefit of it Okay, I tried to make one simple ecosystem or what 5g edge is so far If I miss one of them, please forgive me I tried to come up with the open source project that I'd aligned to the 5g edge As you can see the center is Focused on the CNI CRI and CSI which I've studied today Also, I forgot to add the the the mesh Which is also key for 5g edge, of course these four projects open source I'm going to discuss about them. Yes as a summary because of the time and Above that as we can see there's Argo amazing ICICD Some of them the numbering well now it's been I care that I'm using New few is an open source project is having a lot of attention these days in previous a cube com They have a lot of use cases which is somewhat quite interesting and on up with just a standard to set How it is going to be deployed on top of that of course we have application as I said for us We are developers interators 5g is just not on just an application Of course, it's complex But at the end of the day is just one application and for we have these two option with our open source We are mainly used for any for many international conference. So I try to put them in this presentation the table on the left it's Summary and the comparison between these four projects for cube edge as you can see Slightly advantage is assigned to the k2s And while the cube edge it's having some sort of like a complexity Some not available documentation personally and this is just taken from one research this year personally I didn't find any issue with the kube edge. In fact, I think kube edge and k2s are so far one of the strongest one to make at least R&D or testing test labs But pretty much this is a summary of what we can see for the current top Kubernetes open source projects Okay, so For the demo section I Tried to be as simple as possible. I trying to not show or not focus on the Productive both scaling first because it will demand to have one machine learning and one algorithm to do that Prediction instead. I just simple use one cluster We want master now tree worker notes Public IP and the reason why is that because I'm using or I used a double S Services and specifically one easy two notes, which is just brittle machines The IoT devices in this case are IP cameras, which is a basic solution for just streaming or data streaming Of course, we have permit use for an L a key for monitoring and data and observation The notes a specification in this demo. I try to use a 32 cores 100 GB RAM 500 GB storage and the 233 IOPS When I start with my colleague about this scenario this also that always a question. Why do you pick those values? Well, there are many reasons of course One of the key ones is that as a community builder I have free I Have some points to use so I didn't I didn't have to pay for this So I just want to try how how robust are the the AWS In fact, this let's say flavor is used for the actual 5d core that it was offered to the public So I just want to try The repo has all the the load test the load tester code As I mentioned before in this presentation for the sake of the time I'm not going to do any video demo guys just going to go straight to the summary Please go through these deeper. You have repo One and I must mention this I Basically found one amazing low test repo and fork it and modify it So smart real smart guys make this low test So it's not for me to revamp the wheel. It's just to use and that's right to apply for my use case these Six test cases KPIs is what I consider relevant if we focus on the It at CD or the database consumption if you can see we have a cluster API how is that how far it is and How fast is the create deployment initially if you can see for all these values Which are in milliseconds the one who is still kind of like give us some? Let's say Some discussion is how much time it takes for the pots to be running Of course, we can deploy easily deploy a cluster. That's quite easy but if you have to scale and considering the curing limitation we have for the at CD of course that four projects Have some struggles still have some struggles These four projects are used different CNI CSI and of course we have a lot of customization for a second of these For this demo, I guess I use default a CNI default configuration As you can see QBH has the highest response Which doesn't imply that it's not the the best option One of the the ones that was kind of steady along the the test which by the way, which is it was just a scale out from 10 replica to 100 replicas So is it just one scale out and just I was trying to Monitoring how fast are for the pots to be running because this imply I have 10,000 IP cameras. I have many cameras. So they are increasing the traffic loads rapidly So what I'm expect I'm expect my cluster to get roughly used to this increase and adjust and Also increase the capacity That's pretty much the demo that perform and Have time Okay, take a waste as I said Five you call the edge have many advantages and many challenges one of them and I present today are two the HCD dependency the HCD let's say Let's see the limitation when we're trying to be too strong consistency in that the store There are many options that are many use cases that doesn't require it to be strongly consistency Also, the post strategy is to one of the key topics are still a lot of research a lot of tests a lot of failure and tests doing recently to for the fourth strategy is Trying we're trying to Get out of the basic Of scaling and yes, I use a more predictive one Which was a customer usually requires at the very end I talk about the current edge Q and the projects designed and There's show some speed when we are Starting up when we are deploying the cluster, but still there are some observation or there are still not Where's it seems some latencies when we're trying to actual get the pot's running Which is it when the application is is already are ready for for taking traffic? that's pretty much my my demo for today and with that, thank you so much for your time and Thank you. Oh if you anyone have any questions or any comments, please It's almost five thirty. Yeah, people are tired If not, um, you can connect with me with this in my blog that I have I try to upload a double S and Of course, I do and to bonus content, please you can contact me and get in touch with Twitter or linger in It's fine Yeah, thank you