 My name is Andrew. We are talking about our one of our project adopting the heterogeneous computing resources in our data center which is embracing the arm as a flavor. Yeah, this agenda. Let me give you a quick introduction of our company and the scale. Yeah, we are a number of mobile messenger company in Korea. There's a lot of people using it actually and we have a web portal, e-commerce, mobility banking and other system and services. The open-stakes scale, yeah. We have a 150,000 instances in our data centers and 70% of them is for the Kubernetes already. Actually our services, no, 95% of services transform into the container one. Yeah, so container is a major workaround unit and 600 petabyte volumes are quite large and we have five with six regions. Yeah, in our open-stakes regions. Yeah, and for the scale we developed our model. Actually I give a presentation at Vancouver Summit 2015 and the other storage drivers we made. Maybe we can share this story the next session or this summit. Yeah, my cloud vision is a simple and effective cloud. For the arm service, we made OKA like ready for the global and extra computing. There was the mission or there was the purpose. Yeah, motivation and other things. My team leader, Kihun, will have details about that. Thank you. Thank you, Andrew. My name is Kihun Jung. I'm from also Kakao Corporation and then today are we going to give you about why we start this project and then what's the baseline and then what is the motivations and what is the approaches and then, okay, let's move on the motivations. Motivation number one. So we have, you know, as you can see the screen, we have some bunch of type of flavors. So we put, you know, really recently overall examination for the disc players workload and then we found that we have low CPU utilizations and the high memory consumption phenomenon. As you can see, right side of the figure, so most of the flavor types are underutilized over CPU. However, the memory is quite well utilizes. So what is the reason for the kind of phenomenon? The Kakao service workload is the reason. That's why we are mainly offered mobile service. It means small memory footprint but large traffic. So they are using a relatively many memory but small at CPU and then also our service are mostly long on top of the Kubernetes so that our workload is mainly many but highly distributed over the cloud. That's why we have the kind of phenomenon. So that's our motivation number one. And then second, we have also big needs for the transforming and then inversion very new hardware architecture. For, we have two perspectives for needs. First one is technically and secondly strategically in the edges. First of all, for the technical, we want to use our resources so much efficiency and then we want to also we want low latency IO and then as you know, as a cloud provider, we don't want customer under a noise neighbor situation so we don't want to also don't we don't want to avoid noise neighbor situations and then also we want more fast and effective cloud through the host-based installation and then for strategically, we want seamless introduction with various type of hardware and then we want and then we also saving the legs spaces because our IDC or data centers are now suffering with lack of spacing so we want saving our legs spaces and then due to the kind of situation, we want to also build and running and operating our data center with cost efficiency and then we need new hardware architecture. This is all for our new data center so that's why all the kind of more invention we keep off, we think like that we need something like kind of things are for prepare our future. First, we want to use highly converted server with lower power consumption and then to enable host-based acceleration, we maybe need some kind of separate data processing units like DPU and then we want also white box switches and then finally, we want specialized server over appliance for our specific work load so we have decided to use ARM. So ARM is as you know well, it has many covers and then however tutorial deduced this instruction type so it has low power consumptions. Nowadays, ARM has hand-full two chains and ecosystem as same as x86 so it's very easy to use and easy to programming as an SOC and then recently it has also recently it has been proven by real use cases such as AWS or Apple some so that's the time to we accept new type of architecture and then we it cannot be achieved along with inside of Kakao because we also we need partner company to join this project because our Kakao has a little bit leg of skill to make server or build architecture build hardware architecture yet so we have choose access labs access lab is South Korea first ARM server manufacturers it has this company has top-to-bottom solutions and then it also has production experience so that's the right partner to join this project. So we are a kickoff joint project with two big topic. First one is find feasible workload for ARM microserver using Kubernetes and then the second one is ARM based cloud resource acceleration hardware so all of this story how we start challengeable project in our company. Okay I will move on the next sub some slices we are going to show give you talk about our recent achieve in this project first one first project is research with researching feasible workload so our environment looks like I want to be looks like like as you can see on the screen this is what we want to achieve in high-level architecture through the project number one. If you we have a workload and then the user when you user submit workload to our Kubernetes as a service in our ARM based microserver participating as a braver to open stack based cloud and then Kubernetes as a service use that kind of flavor as a worker nerd. Each server powered with very relatively very slow but very low power consumption so it will be quite challengeable how how what kind of workload fit for this kind of situation so and throughout through this project we have some we have derived four kind of discussion and questions first what kind of workload types are feasible for ARM secondly does ARM really offer affordable performance per watt than x86 third how can we introduce new architecture type into open-stack cloud seamlessly finally is our in-front system ready to adopt new architecture so let's move on this culture three and four left science process are existing existing process there is not our responsibility but but it is in our input team so right side our researching target properties ARM is totally due for in our company and then our in-front system as well so we have well well configured and performing existing adoption process we have a provider network centralized OS provisioning system and homegrown with tools and security compliance tools so we want to integrate with our new hardware type very and then existing adoption process are mainly concentrate for and build for the x86 so we want to introduce our new hardware type as an ARM and so our new target has BMC-based fixed DHCP and then also BMC-based internal OS image system and then it has not IPMI not ready that fish not open BMC it has on their own customer management system so we want to overcome this kind of challenging so we some picture are not acceptable or migratable to our legacy process so we want we find out figure out how we integrate with our system this is ironic way we build some Nova computer drivers and then it drivers translate our target system to into the open stack cloud as a flavor so through that our ARM systems are participating as a flavor into the open stack cloud without any challenging and any legacy broken system make make legacy broke broke so and then shortly do we have any problem with the Kubernetes thanks for the communities and supporting we just not not ready for the arm yet in kaka kaka only so every code and tools are already built on ARM and we just build and run and then everything is fine and then after that kind of our solutions we can change our target to into hardware as a flavor the DHCP we use provider network over the DHCP and then we can input to external image to into the BMC and then all of these thankfully through the Nova compute and the ironic driver so we make successfully introduce our new type of hardware architecture into our cloud and then number one and number two discussions are feasible workload so due to the speciality of properties of ARM we we can think like that feasible workload maybe this two rebutt and then parallel parallel execution so then same workload work throughput but low power consumption maybe expectation this is our short benchmark result from that if we have some short leaving HTTP connection benchmark and then encoding test as you can see on the screen the throughput over throughput or in the request per second and then including time same with we match same number of beat x86 and then we found that the the power consumption is eventually very lower than x86 style so when we when we properly scheduling delta over the arm microserver it will promise us same throughput but lower power consumption is it is very quick summary of our project number one result we're learning out the time so we we are going to very quickly overlook our number two ARM based hardware acceleration so we have we are using now very our unique host routing model we are also offering SRU we pass through but somehow we have we are all actually worse so we don't have L2 layer for the TSP or some we can not embracing L2 layers property so we are we are doing at inside of corner so this this is quite fit for over performance so we need something like that something like helping us to make these things in inside of hardware so we want to we want to build new our we want to build new type of hardware like smart nick so we are now build new type of hardware I'm associated based in smart nick and then this is the chip from the NXP and then it is already DPDK radio on the SOC and then it provides some hardware acceleration features so we can achieve this kind of thing through the this kind of things and then we can make it more affordable cloud and high performance cloud through this kind of things and then also we were preparing the ARM based white box but the purpose over some usage or use cases are still figuring out so okay let's move on let's summary of our presentation in this project we we learned that kind of lessons first work for work process and cultures new hardware type is always doubtful so that we need to move step by step there in kind of and then tech tech is not only the solution so we you we are very okay we are so carefully make a solution with this kind of processes and then RAM is feasible for with smart enough workflow engine workload can be sliced and distributed beautiful so using many cores in very short time means that eventually the result is performance per watt winner and then in perspective of user for user several architecture is already hidden by open stack abstraction layer yeah that is so that kind of flavor so each flavor computing power power is critical factor to choose the to their usage so if the performance of ARM is relatively same as x86 no reason to not no reason to not use using ARM so lastly we are some very special situation right now so we have we are under suffering with chip shorties the issue is bloody effect so that if you guys want to implement new hardware implementations should be carefully and plan with this kind of situation okay thank you for concentration to our session and if you if you have a question okay all of the time so the question and answer will be in privately so thank you for listening thank you so