 Hi. Is everyone doing fine? My name is Jira Chung. We're from Orchestral, a cloud stop for a company based in South Korea. Today, I'll be giving you about a 15-minute presentation about optimizing service continuity in a multi-cloud data center. So to give you a background pretty much, I'll be going through the table of contents later on. So this will be the table of contents. First, we'll be going through the background to give you a better context to understand the challenges that we faced and why and how we tackled those challenges. Second, we'll be talking about the challenges that we actually faced, but specifically in two areas. So in Korea, a lot of customers are adopting the multi-hiver cloud model. And they're also utilizing AI cloud to provide their services as well. And secondly, because they're adopting a multi-cloud model, they're adopting VMware and OpenStack within their data centers. So they would like to utilize both of their features in both pools. And then thirdly, we'll address how we actually tackle these issues. For regarding the AI cloud, we are going to talk about SmartNICs and OVS offloading. And then for regarding VMware's DRS functions and mimicking those functions within the OpenStack pool, we'll be talking about how we were able to mimic them and provide them as a service through our portal. And lastly, we're going to talk about future challenges that we're also facing, and specifically about active, active data center architectures. So before we get into deep about the presentation, I'd like to give you a brief introduction about who we are to give you a better understanding of the context. So we're a cloud software development company based in Seoul, South Korea. We actually started in 2018 as an AI ops company. And these are the full stack of cloud solutions that we provide. There's cloud solutions that we provide to deploy private cloud environments within the customer's data center. So we don't provide cloud services directly to end users, or do we have data centers? But with clients that have data services that are willing to go through digital transformation and cloud adoption, we provide our cloud suit. So from the bottom, we provide the cloud infrastructure level. So OpenStack based IS infrastructure, we provide that. And then Kubernetes native. And then in the middle, you see an orchestra CMP, which is a cloud management platform. Well, our cloud management platform does is integrate multi-hybrid cloud models. So whether you're using VMware, Red Hat, OpenStack, or our OpenStack, or even in the past layer, TimeZero, OpenShift, our CMP provides an integrated management for that as well. We also provide support AWS and GCP. And under the top layer, we provide data ops, DevOps, ML ops, and AI ops as well. So this is just a brief context that we provided to give you a better understanding of the next slide. And so in Korea, there's a big trend of adopting multi-hybrid cloud models for when they're adopting the cloud. And these data centers specifically are very big. They aren't just a single data center, but they're located in different regions. And they require us to connect them in different ways so that they're actually all utilized in a way together. And two specific needs and challenges are going to be talked about and mentioned today in today's presentation. And the first need and challenge we faced was due to the increased need for AI cloud services, they were requiring us to enhance network capabilities as well as CPU performance. And the second need and challenge we're going to talk about is mimicking VMware's DRS function in the OpenStack pool. So to give you a brief introduction about the very summarized architecture of one of our clients' data center, as you can see, they have adopted the OpenStack VMware for their private cloud. They've also adopted the HPC for the AI cloud, so high performance computing pool. And they also adopted the public cloud. And all of these cloud environments are being integrated management by our CNP on the top. So first thing that I wanted to explain that is a good context for you to know is that they want to provide the services in a unified workflow. So for VMware's DRS services, they also require us to provide similar services in the OpenStack pool. So it was necessary for us to also provide those services within our portal as well. So with that background, let's get into the first challenges that we faced. So the first challenge that I'll be talking about is enhancing network and CPU capability for AI cloud. And there's many ways for you to tackle this issue, but we'll be talking specifically about how we integrate a smartNICS to enhance them. And traditional in network setups, as you can see in the diagram below, there's TAP devices in between. And these TAP devices with an environment where you require high workloads data to process, it could cause TAP device overheads, causing bottlenecks in trafficking and bottlenecks for CPU performances as well. So because of that, there are needs for us to tackle this issue, to decrease TAP device overheads. So why smartNIC and what are smartNICS? And before we dive in, kind of giving you talking to you about what smartNICS are and their benefits kind of probably give you a better context of understanding why we implement the smartNICS. So smartNICS are able to enhance both network and CPU capability by offloading them. They provide very various functions. So not just only network functions, but also other functions as well. But, and it is a very common way of solving CPU and network performance enhancement. And as cloud technology is maturing, infrastructure advancement is moving towards away from traditional CPU centered cloud infrastructure. So it needs more advancements in different various ways. And in this context, especially in AI cloud, data centers benefit when acting as a unit of a CPU. And to do so, you need to be able to allow the CPUs and networks to be functioned in a different way in a, so that they're more optimized in their best way capabilities as possible. And to just address these demands, a lot of companies are equipping data centers with specialized hardware accelerators such as GPUs and ASICs as well. But today we'll be specifically talking about SmartNICs. So SmartNICs is, as many of you already know, are advanced network interface cards that offer dedicated hardware acceleration and offloading capabilities for network related tasks. They handle tasks such as packet processing, encryption, decryptions and all sorts of things. They're very complicated to touch with, but they're also very useful. But today we're gonna talk about offloading these network tasks on the CPUs to optimize them. So this is a basic diagram and architecture of how we utilized it. So SmartNICs operate like a server inside of another server, allowing virtual machines to communicate directly without the NICs to being interrupted by the interfaces in between. So they decrease the tap device overheads and the interfaces between the kernel layer. And to enable these functions, we customized the FPGA code within the SmartNIC to provide the obvious offloading and capability and the obvious DPDK capability. So to talk about the obvious DPDK, obvious DPDK as you already know, it allows host to path through. So direct connections between virtual machines and NICs. But even without the SmartNICs, if direct connection is made through virtual machines and the hardware and the hypervisor level, there's still great overloads within the hypervisor CPU. So to optimize the hypervisor CPU function and performance and the overall data center as a CPU performance, you require a SmartNIC so that you kind of offload the hypervisor CPU onto the SmartNIC. So the SmartNIC is actually functioning as a network optimization aspect. So that's how we provided optimization for VM path through and direct connection with the underlying physical level, late layer. And then second is the configuration of OVS to use SmartNICs to provide neutrons functionality. So neutrons as you already may know, provides various network functions from East-West traffic as well as load balancing firewall as well. But so we offloaded those tasks to be performed on the SmartNIC directly. So the SmartNIC are performing the load balancing firewalling, QOS management, TLS security as well. So this kind of decreases the overall workload and burden of the CPU to the SmartNIC so that the CPU is more functioning to more focused in the hypervisor optimizing level. So in a similar context, if you look at this slide, we are talking about East-West traffic and with the rise of cloud computing, distributed computing, cloud native applications and especially AI cloud, East-West traffic has become a very crucial point of challenges that you need to tackle. And East-West traffic in AI cloud especially, it requires high volume of data to be processed, to be transferred and leveraging SmartNICs to offload these capabilities so that the hypervisor CPU isn't overloaded, isn't burdened too much, optimizes the AI cloud in this sense. And yeah, so in this way, we tackled the network enhancement capability and this is basically a data that we collected from, is between an environment without the SmartNICs implemented and an environment with the SmartNIC implemented and you can see that the network traffic functionality and capability has increased by about 70%. And so with this, I'll just kind of finish up and wrap up about the first challenge that we challenged. And secondly, I want to talk about the, mimicking the VMware's DRS function in the OpenStack pool. So DRS, to give you a brief introduction about what it is, is a distributed resource scheduler and it is a feature of VMware within vSphere to optimize resource utilization and workload distribution in a cluster of virtualized host. So it's pretty much a package solution service of load balancing, server group clustering and cluster management. So about affinity group, you could anti-affinity group as well. So with this, you also provide optimization for the administrators so that they could improve resource efficiency, overall stability and scalability as well. So this is how we tackled this issue. We weren't able to mimic 100% of the VMware's DRS function itself, but we were able to mimic the functions to this level so that it could be provided as a service through our portal. So what we did was we combined with automation tools, Octavia service server, OpenStack server group and DRBD together. So we combined them in the back and we used automation tools to automate this whole process. So to give you an overview of the process, the first step is when you provision a virtual machine, our OpenStack server group them together so within different hypervisor when you're making a virtual machine, you're provisioning it with a certain image from glance and you're provisioning applications as well. But it's provisioned in an active standby way. So even if when a virtual machine from hypervisor one would go down, the virtual machine on hypervisor two would kind of go live. But the real important thing is how are you going to synchronize the data? How are you going to provide it as an HA high availability architecture? So to do so, we linked these two virtual machines through Octavia and the VIP. So load balancing is pretty common and basic feature. So I'll just kind of go over that. And then lastly, we synchronized the data box between the two hypervisors and virtual machines so that within the same server group so that they're able to provide high availability. So even if one virtual machine fails due to whatever reason, you already have a synchronized data within the other virtual machine within the same server group as clustered by the open stack server group. So by providing these services, we're able to mimic the same benefits that VM or DRS also provides. So efficient workload distribution, high availability and failover, and customizable policies about affinity groups. And lastly, I want to talk about the challenges that we're facing right now. So we not only provide cloud solutions, but we also provide technical consultancy as well. So this is an architecture that we're drawing for our one of our clients right now to create an active, active data center environment. So there's four main things that you should take into consideration when you're designing and architecting a active, active data center for multiple regions. So the first thing would be GSLB. So when a client is requesting a certain cloud service, you want to be able to provide the cloud service in the most optimal location and region possible. So through this GSLB, you're distributing the request to the most optimal location and optimal data center. And second is we're actually partnered with Cisco in Korea. So Cisco ACI VX line, we're going to use their L2 device so that we're going to connect the two data centers together. And then third is GWDM. So what you want to do is because you have to have large amounts of data to be transferred between data centers, you need to be able to expand and manage the bandwidth and the network connection. And lastly, you want to have a storage that is able to support active, active environments. So not storage is one of the, you know, storage is that we find that is very flexible when you're utilizing for active, active environments so that you don't have to, you know, it's more continuous. Service continuity is very crucial. So you don't want to have any downtime when you're transferring the data between the data centers. So the active, active data center model isn't a challenge that we have actually solved right now, but this is just one of the approaches that we're taking. And hopefully it provides some insights. If you do have any insights or opinions about this, please feel free to visit us at our booth. We're more than open to talk to you about this. And if you have any questions regarding the contents that I've talked about before, please also feel free to visit us at our booth. I'm not, I'm actually in bargaining sales, so I'm not super like in too deep with tech. So if you visit us, our director and head of our technology would be there to answer your questions. So yeah, and with this, I'll end my session. Thank you for listening. If you have any questions, I think my tech guys here, if you have any questions that maybe we can answer. Okay, like if you're interested in anything, then please come by and stop by our booth. Thank you once again.