 Thank you for taking the time to listen to me. So Jim started out by going back in time and say hey We were the Linux foundation. We did Linux and he showed a slide from like six seven years ago I have a slide from five years ago as well five years ago we showed really how As much as containers were hey, we're gonna make virtualization obsolete virtualization is for floppy disks The technology behind virtualization and containers was sort of the we were rethinking that space and we launched your containers a little kind of containers and I Have one like sort of flashback slide where we could talk about what happened Containers came up containers were what isolation and containers were about how we deploy software No containers change the world containers change How we deploy software how we develop software? completely At the same time security was a question. How do we secure containers? How do we isolate? How do we do? interesting things there and In one of the staff meetings at work where some some guy in the staff said hey, well Fertilization is that we should stop this whole virtualization thing and we'll go do something else Because it's too heavy too hard like I Asked myself and we asked it out loud. How hard can it be right? The most dangerous word to computer science. How hard can it be to actually? We've reimagined virtualization and we did it with clear containers in the beginning all we showed was we can do a very lightweight isolation Oh, and by the way, we can glue it into Docker Although Docker was kind of at the time not mature enough to have plug-in architecture So it was hard to make it work over the over the last few years We went from something that was really hard to use. It was an interesting concept who a Dockerman modular OCI showed that Containers are an ecosystem not an implementation and there's multiple ways of making back ends as multiple ways of doing containers and Wasn't just us the hyper V company Did the same thing around the same time and at some point we all decided to sort of join hands and pick the best of technology of pieces Just from the previous speaker you have multiple innovation points And at some point you make a step together. You combine two new things together and Then Kubernetes happened and Kubernetes was basically saying yeah containers are great But here's how you use them. Here's how you deploy them. How you put them into production use and it was another whole bunch of work to make sure Fertilization containers all that work together very well and In the last few years We're now at the point where the Cata container technology is mature People are producing their production our friends as soon as they started shipping it in their OS While gaming people are using it for gaming in the cloud as a way of having high density, but still secure kind of gaming Sort of it has arrived. Okay, so why am I saying this? Why am I talking about the past? Well One of the things that changed was really Okay, so virtualization technology isn't about Enterprise VMs that emulate the floppy to the expense control system from the 1980s can still expense process your expenses. It's really about It's lighter. It's about isolation. It's about that the basic of the technology and It's basically a spectrum the problems were transition technology can solve for you are not just virtualization as an enterprise technology and Earlier this year that the firecraft Amazon guys launch firecracker and they showed hey lightweight virtualization technology is great for fast Not even containers never mind containers containers are big and heavy. Remember virtualization was big and heavy containers are big and happy Let's do fast and They use the same kind of technology for isolating fast from different customers from each other and it really One of the things our group realized is and we talked to many of our customers is well virtualization is Changing what people need from their isolation is changing So what what do what do people want? We talked to our customers We talked to the community folks. We talked to a lot of people and sort of five themes came out of that The first thing was ever ever we estimate lightweight Well, it means many different things to many people but lightweight was something that came back over and over again fast Well, no everybody was fast But what does fast mean? The more tangible one was density people want to use them. They're the same hardware for more and more more work If Amazon would run one fast or server, they wouldn't be in business. They want to run hundreds or thousands of fast jobs on the server This is to be quick The world where you install a server install a VM leave it running for three months and ensure it down has gone The files job runs five ten milliseconds and it's going again if it takes you three minutes to start Doesn't matter you don't you don't exist and security is almost not negotiable Jim earlier launch secure computing, but security as a concept is Every customer we talked to say security as a baseline you cannot negotiate security You can negotiate density you can negotiate startup. You can negotiate lightweight security is a bar. It's just a baseline Okay, so we really and we went back to the same customer said, okay So you mean lightweight, but you're for fast. What does lightweight mean? two megabytes And I went to the guy who runs the clouds and the clouds for station layer for for a big customer What does lightweight mean? Well to get two gigabytes so some of these things are proportional to the problem space they're in and That means there's not really one size fits all At startup time same thing fat the fast guy say yet my whole job runs ten milliseconds a fast startup is one millisecond or less But if you run running your enterprise VM in a big cloud provider and the thing runs for three months few seconds who cares so All of these things are proportional One of the challenges you always have is okay, so from a technology perspective, what do you do? you have this wide range of requirements from very rich requirements with torrents with for for for lightweight for density versus to very fast and one things we started working with with a bunch of our community partners is Something we call Rust VMM and that is basically rather. I'm building one hydrovisor or one virtual session stack for everything build a set of components that you can stack together in any way you want in Order to make it to make it easy to make a domain specific solution that still shares when you have to come on building blocks So the fast guys can use half of this or maybe a third of this But if you want to make a full enterprise hypervisor use all of it or Use different versions of the same block you can make a lightweight version of the device modeling can make a very fat Rich version of the device model and swap them out So rather than having one Virtualization block you make we build a set of building blocks that are replaceable that are optional They can very fast decompose into a domain solution all the way from fast or lighter to all the way to enterprise or richer Without having to reinvent the wheels or the things that you can't share Now how do you build the set of building blocks like that without actually getting lost? So one of the things one of my team did is okay, so let's also build an actual implementation of this that is relatively rich So that we can verify all of it works together This is what we call the cloud hypervisor. It's a full hypervisor for a virtual machine setup It's it's an early phases So don't expect to be competitive to be in the production tomorrow But we're using this to prove that all the blocks for building of components can be used together in a rich setup But also we have configuration of this in our last on that And it turns out you make a hypervisor. You need more than just these building blocks You need to add device back ends to it. What what kind of stories do you use? How does your networking work? But if you don't prove out the set of blocks you have no idea if the alpha work together So we have to build blocks that put them together in addition to the light with fast of the Container site. This is also sort of the other side of the spectrum So when we say reimagining virtualization it is really about being Composable being domain specific while sharing as many blocks as you can from a common building block set Okay, so can we solve virtualization? Let's move on My team is very busy with this and we're spending a lot of time making sure we get this right because virtualization and then The OS around it is a building block for everything else you do in your computer Okay, so The second thing that Jim Jim allure by spam my time on is okay. What's what else is there? And Jim kind of implied this earlier as the Linux foundation went from the kernel all the way to networking to higher up no gs higher up in the stack So did we in terms of how we look at software? So far isn't just a kernel so far isn't just a virtualization layer so far isn't just Some library gipsy and maybe a compiler here or there so far is all the way into an anti application. I'm back and In my day in my spare time when I can actually do coding like an hour or two a day. Yes. Yeah, I do code I Look at like looking at performance because performance is a place where you can look at things in a different way and get a different outcome So when we looked at performance for a bunch of basic machine learning and all operations it turns out that Performance is not a problem. You solve in one spot on the stack Performance is something you have to solve in every layer first all work together and it's a non-linear effort We found okay. We saw performance in matrix multiplies meant a few Saturday afternoons matrix working on matrix multiply It was great. I got a six X performance improvement. That's a great Saturday afternoon, right? It turns out if you deploy the same algorithm in specific cloud provider You wouldn't get any benefit because they didn't actually pass through some of the basic data. You need to get that performance improvement Oh, and by the way, then the layer on top the machine learning layer had a matrix that which was wasn't nicely 64 by 64 It was 65 by 63 and Not a nice power of two So you have to change the machine laying layer above it to say can we make the matrix at least sort of multiple of eight To get nice performance. So performance turns out to be something you have to do all the way up and down That makes it an interesting challenge and an open source. We have the source code for every single layer of the stack So that became a Sunday afternoon project Can we do the rest of the stack as well? and the answer is yes, we can and At work we realized that this is a hard problem a Lot of people in the open source world have their own sort of project and they know kind of other that happens around it But not a lot of people really look at what's up and down and from an end user. It's even harder If can you imagine being a system administrator and having to deal with eight layers of the stack and making sure they all work together That's a hard problem So one of the things we started doing at Intel is okay Can we just publish one of these make sure it works? Well, and then have people and even if they don't use it They can at least see how it works together and they can compare their version of this with what what could be And if they come as a way of learning from it, hey, if you want to use it you can download it It's great. If you don't want to use it. That's fine, too but if your performance is not as good as what we show you at least steal the performance bits from it right, we want to show you what can be done and how it is done and This is almost integration exercise because it sounds boring, but Putting it all together in a way that works together is actually harder than it sounds So we did machine learning that worked great. There's like four or five people devops team measure every day and just keep tweaking at it We ended up doing something around data analytics data analytics is a even bigger problem and Spark is a Hadoop a great project, but getting sparks set up correctly using all the layers in the stack is actually an incredibly complicated challenge Because you have Virtualization you have the kernel you have NUMA you have network you have all those components if you get one on them Not optimal the whole performance collapses. So we decided let's do one for spark And we did that and that was fine. We go For machine learning we go to 12x performance increased by doing things right now I don't like giving these numbers because people say yeah, but but at least it's not 5% Right there's real performance on the table doing this, right? Maybe you're in the middle of there and you can only get 2x more that's fine But these kind of things the difference between getting it almost right and completely right is completely non-linear We did that now Machine learning and data analytics are pieces. We are also realized we have to make a much more complicated system This was a this was basically sentiment analysis, which was about okay We have a bunch of text coming in from let's say Twitter feeds and order of sources How do we know a product announcement? I don't know well because it's very typical people think people doing marketing or in the retail and it turns out It's a bunch of those vertical sex I just showed you chain together in a way that actually makes them work together And they have not a layer of complexity making sure they work well together And also this you can also get on our website as a way of hey You saw you all put all these pieces together to get end-to-end performance for your total problem not just for these individual pieces Okay with that Time to get my paycheck back My boss wants me to want you to know that we have an infel booth We have a lot of demos all the things I showed you we have demos in the booth upstairs There's a winter session from my co-workers. We're talking today and tomorrow and Friday. Please visit them Come talk to us at the booth and thank you