Red Hat OpenShift OpenShift Coffee Break | Red Hat Livestreaming, 15 Jun 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: OpenShift Coffee Break: OpenShift Sandboxed Containers

Description

Get your espresso ready for the EMEA OpenShift Coffee Break as we welcome our special guest Christophe de Dinechin, Software Engineer at Red Ha to talk about OpenShift Sandboxed Containers!
This functionality is based on the Kata Containers open source project and provides an Open Container Initiative (OCI)-compliant container runtime using lightweight virtual machines running workloads.

Twitch: https://red.ht/twitch

A

A

Good morning, everyone.

B

A

And welcome to today, official coffee break uh with my dear colleague, corgis jafar, one in japan, hello, everyone.

B

I've had you history.

A

Hello, hello, christoph, is our guest. Today I will be having the opportunity to present itself, so I will not add anything else. Please talk. Could you introduce yourself.

B

A

C

Obviously yeah good morning, let's start.

B

Together have a seat before.

A

B

Always we are open to.

A

All kinds of coffee, so italian, coffee, american coffee, turkish, coffee, whatever you.

C

A

Please kiss stuff.

C

So I like my coffee, cold and with cream, so I'm really bizarre, uh so I'm christopher dinosaur and I'm working on a product called open shift. Sandbox containers that we are going to discuss today. uh That's the openshift implementation of an upstream project called catacontainers.

C

uh I hope that today, we'll also have a chance to discuss about some further applications down the road like confidential containers and otherwise I've been working in the past in virtualization, 3d, graphics, compilers and operating systems. So that's essentially my background.

C

Not all of this was by the way, unfortunately,.

A

That's okay, we'll all have some not open, shows.

B

Behind us, it's.

A

C

B

C

The good side, I've always done open source since a very long time, but sometimes I was doing open source on the side and then had my so, for instance, in a very a very long time ago, I did the first part of emacs on mac os 10 back when it was called rhapsody. I did the first graphical part.

A

So kata content, something related to any martial arts or.

C

Yeah, so actually you have to be careful about what you're talking about. I saw someone on twitter was saying that he read about cafe containers, so kafka containers is a very different thing. It's when you know. If your containers are too fast, then you have this option of sending a form in triplicate by fax, to your cluster operator, and then he will start your container within six months. So that's a different product. Kela containers is the idea of running containers inside virtual machines um to provide a bit additional isolation.

C

Let you, for instance, configure your channel to your taste and run things uh as roots without having risk for your host and this kind of things. So this is really um uh a way to exploit lightweight virtual machines. It's a bit counter intuitive, because if you remember initially, containers were seen as oh that's to avoid virtual machines, because virtual machines are heavy weights.

B

Yeah and but why in isolation as well,.

C

Yeah, so so the problem with with containers is that um they run in the host operating system and exploit its resources, exploits um whatever isolation features it gives.

C

But then people start thinking, oh yeah, but I can bypass that, uh for instance, intentionally in the case of uh privileged containers, which is used often, for instance, to install software on the host other things like that, and and sometimes do you have no other choice. For example, if you want to access a physical device, like I don't know, an si of unique network interface card or all these kind of things, sometimes you need at least part of the configuration to be done using privileged software and so with color containers.

C

The idea is that you can run this kind of software in its own virtual machine and that's now feasible easily thanks to moore's law, because something that 10 years ago made little sense. Today we have more resources and so on. So that's something that we're going to fold and now you can, for instance, attach a physical device to your vm and and let the software inside control that physical device directly, while still having um isolation enforced by the I o, mmu and so on.

C

So this gives you additional possibilities for for what you can do with containers.

B

C

B

To to clarify uh how this is uh executed, so um you said it allows you to run clouds in vms. But uh do you mean it's going to run the pod as a vm is? Is there like a one-to-one um mapping or for each pod correct.

C

So exactly so so, essentially the units uh corresponding to the the vm is the pod and the the rest of the deployment is relatively similar.

C

We use some features like virta ufs, to expose the host file system to that virtual machine so that, for instance, you can download your container image on the host and do um you know, have the overlays, etc and and then that's transferred to the guest and the guest sees the file system that part of the file system of the host in its own file system.

C

uh You can also attach things like block devices. You can have other options for how you connect to your your storage, um the networking same thing. You have uh several options, but you can go through um turntab devices, etc. So there are various ways to to connect to the enclosing networking.

C

But then you have your own network and you can have additional capabilities, for instance, compared to what would be safe in the host.

B

A

Yeah there is always, as you mentioned before, there is always there is been a great.

C

A

uh On the topic containers versus virtual machine now we are bringing the concept of virtual machine on the ship. Can you uh explain us better uh when actually before, which is the main difference between having vitro machine and sandbox container or sorry container machine in our sheets unboxing container, and um why why I should use one of these.

B

Yes and maybe to expand on that so, for example, the difference between containers and uh and using something like which is also the vm inside, I mean a pod as a vm, but what's like the difference, and why is there this new initiative.

C

So um that's a that's a very good question.

C

The idea of uh keller containers is really to have virtual machines in the context of the container's ecosystem, so in other words, you're really running containers, whereas in the case of cubot you're really running virtual machines and you can run, for instance, windows, virtual machines, but you're orchestrating them with the openshift kubernetes execution, environment, and so that's the main difference is that in one case, you're really running containers. That's the case for color containers and you're.

C

Just you happen to run them in virtual machines, but that's sort of an implementation detail, whereas in the case of kubevert you're really interested in running um a virtual machine as a virtual machine and to use its benefit.

C

So if, uh if, if it's okay for me to share a screen, I have a node slide dash.

C

um I I'm not sure if it's public, so in the sense, I'm not sure I can share the link of it. But let me try to just.

B

If you share the screen, you should be able to do it understand better. Okay,.

C

So does it work, yes,.

A

Perfect, yes, yes,.

C

So this slide deck is a bit outdated um because it dates back from a presentation I gave at def conf in 2021 uh defront.cz, so um but that's really um an overview of of what we are doing here and so essentially, what I just told you is is what is shown on this slide here, where we are trying to really benefit from the container ecosystem, and that's mean that means a new architecture for application that is scalable, it's small, modern, etc.

C

You can provision components very easily. You can do software composition, you can you have other benefits of dynamic and automated orchestration in a recent coffee break about takedown. There was uh this idea that your you can really automate practically anything with openshift and I think that's very true, but at the same time we want to get some of the benefits of virtual machines, and that includes kind of tuning configuration uh hardware, enforced isolation with things like the immu et cetera, and that includes isolated access to devices and in the future. That's the confidential containers approach.

C

We can benefit from things like memory, encryption where you can completely hide your guest memory from the host. So that's essentially uh the kind of benefits that we can get from doing things this way.

C

A

Forgot to answer.

C

One part of the question: I'm sorry uh and that's the question about who uh benefits from this, and so some of the markets that um we see having whether there is subtraction for this include either markets where you want to benefit from the isolation for hardware, containment, and that includes, for instance, uh the telco ecosystem, uh this kind of of environments or uh when you want to have isolation for legal or various reasons, and that includes banking and this kind of environment. So these are the two big finance or these kind of things.

C

These are the two uh areas where we see traction first.

C

I suspect this would yeah. So please go ahead, no, no! No! No! No! No! I suspect this would become more commonplace in the future as the technology matures and the relative cost of running a pm decreases.

C

So there are many things that you may want to put in vms. Further down the road. That's you know, that's a trend. You see, uh for instance, an open shift for you, and now uh we have these requirements about uh enforcing s linux and forcing values slightly more restrictive policies. There is this trend that we are trying to secure things more over time, and so I suspect that uh in the future, containers will be used more often because of that.

A

Let me see, if I understood well or could we say that kata containers are provided as a sort of another runtime, so you can.

A

B

A

The tools so that we already have an openshift like the github's ones, also image one so on so forth. So from developers perspective, uh you don't lose a lot of features that probably you miss when you.

C

Just virtualization is that exactly so? Okay, so if you um so, if you know a little bit how kubernetes and openshift run inside there is this thing called the runtime that you just mentioned, and that's essentially the part that is in charge of managing the containers while they run, and so uh the historical one is called run c, uh there is one called c run that is more efficient. It's rewritten in c is much smaller.

C

Much faster uses, less resources, so that's likely to become more commonplace in the future and uh and keta has its own runtime.

C

That will take those ci commands and convert that into operations like, for instance, when you say creating a pod, it's going to start a vm and when you say, create a container it's going to forward that request to an agent from running inside of yam in order to create the container there.

C

So, yes, it's a runtime.

A

Okay, cool thanks. Oh.

B

This seems very uh nice in terms of, um I would say, um architecture and the type of features that it can add. But do you have uh maybe a demo that you could show us, so we have a better grasp yeah how things work.

C

Yes, so um this is a cluster here that has uh cadet containers running. Oh sorry, openshift sandbox, containers running, and so maybe I should start with the console, that's probably easier to look at it from a ui perspective, and so the way we actually deploy things is with something called the openshift sandbox containers operator that is going to manage uh the a new runtime class called cata, and so in order to be able to run a workload you're going to simply add the runtime class cata to the world close workload description.

C

So let me shift to.

A

Crystal please sorry for interrupting. Can you please just zoom in just a little bit because, yes, I'm seeing the streaming? Oh okay! Now it's okay, perfect.

C

But the problem yeah I no longer have the menu, so so here I'm in this project demo and I have, for instance, a jenkins deployment here, and you see that the way it looks is, from your point of view, is practically uh the same way you would usually and the yaml okay. Now it's please wait.

C

So uh can I make the text of the year bigger? Yes? Well, if it's bigger, then we don't see much so. Can you still see it here.

B

Yeah, it's fine! It's fine.

C

So if you look at the spec uh the thing, so everything looks pretty much like for anything else, with the difference of this runtime class name cata and that's really the thing that is going to make the difference. um So let me just shut it down and.

B

And so is this something that is added uh automatically by open shift, or do you have to tag it for you, for example, in your deployment uh yaml, or something like that? Yes,.

C

You have to add that to the uh to the yaml file, you can do that either manually or you can have values tools. That would do that for you, you can have an admission controller. You can have various things that that word under specific conditions.

B

C

B

So did you want to remove.

C

C

Why is it complaining? It doesn't exist.

C

That's well it's there, so why is it complaining that it's not.

A

C

Can you go back to that's? A demo effect did that five minutes ago.

A

Are you in the right namespace? Probably yes,.

B

C

I'm in demo I think yeah yeah.

B

And what do you have in your deployment.

C

Well, so, let's, let's, let's check the drink, the trenches deployment itself, um I'm jenkins! Okay!

C

So that's essentially what you have and.

C

Oh, I think it's because I actually saw it. I think it's because I you changed the name.

B

C

No, no yeah, it's uh it's that the one that is running is the one running. That's! Okay! That's right! Sorry! I got confused for two minutes in my file so uh and- and that's precisely what I wanted to show is, if you do a diff between the end skins deployments and the en keynes scada deployments.

C

That's essentially, the difference is the name that I gave to the app so that I can run the two at the same time and this runtime first name, and so in order to start it. I can start it from the console or I can start it from here and that's just so. In that case I want to get a one and that's basically it okay. So there is this uh new warning that appeared in um uh for um openshift 411, uh that's due to the new security context, um but that doesn't prevent that from starting.

C

C

And so it starts really quick yeah. So it's uh it. It takes a couple of seconds more to start in some conditions because we need to run a vm, but that's essentially uh what what you need and then, of course you can on top of that, you can add the services or whatever. So it's just uh and the service itself doesn't need to change so.

B

And so on, on what uh run time does the v? I mean what acts as a host for the vm? Is it the node itself? Does the vm run aside, how does it work.

C

So I think your question is: where does the vm run? Yes, yes, okay, so.

C

So if I look at the nodes uh that are running here- and um so essentially you see that on the roles, this one can run keller because the operator was deployed on it. Okay, um so let me start with the installation itself uh how you you actually deploy that thing. So what you're going to do is so, besides the usual subscription and etc, um you're going to.

C

To install a color config like this.

C

And so uh well, this one is is one that I I take directly from some internal builds, so it has a this additional element to the spec, but essentially you're you're deploying this, and that's going to give you um it's probably better. If we watch it, yeah.

B

Is it available from the operator hub? Yes,.

C

Yes, okay, so let me go to open shift sandbox. So if you just type sandboxed here.

B

Yeah but you have to enable uh show default projects, uh the the switch.

C

C

And you see the operator here, and so you see that this is the one that is installed is from my own source.

B

But there is one.

C

Coming from red hat, that is the official one: okay, okay, um and so uh the.

B

Sorry, sorry does it require anything specific on the node, any specific configuration or it can run on any. uh So it's running on core os.

C

So it's running on standard core os. It doesn't need anything specific except being able to run virtual machines, but practically anything to the cam. Okay and we are working on integrating.

C

With the no discovery operator to be able to uh do more fine-tuning on the kind of machines that you want to deploy this on, okay and so.

B

C

Now you have to put labels, for instance, if you want to restrict it to a specific um set of nodes. When you deploy the operator, you have the ability to say I want to deploy it on this label and then.

B

And does the node need to be like on bare metal, or does it work also, if it's a virtual node with nested virtualization, or something like that? How how does.

C

This so this one is a bit tricky, so we do most of our developments uh with nested, virtualization and actually the cluster. I just showed you uses the nested virtualization, but it's not supported and it's notably not supported by the the values.

C

Providers- and so that means that at the moment you need to have your nodes on bare metal. Okay, uh we are working on something called peer parts. So if you search peer parts, you're going to find some efforts in this area and the idea of peerpods is to use the runtime to start a vm not directly by starting, uh for instance, launching a qmu instance, as is the case today, but instead by requesting the vm from the cloud provider, api.

B

C

Then you connect to it uh over a tunnel or something like that. This is work being this has been actively worked on for clouds from you know, vmware or others, um but it's still working progress.

B

Okay, so it's so so the vm runs on the the the host, like the cloud provider exactly or or the hypervisor itself assigned, and not as in.

C

Yeah yeah, so it's not so it's no longer nested and that means you're you're back in the supported configuration um it's more efficient.

C

uh It's also necessary for a confidential container or similar technologies where nesting is not just not possible. Today,.

C

That's a very good question, though: it's uh at the moment that's one of the weaknesses of the product that we are working actively to to fix. Let's.

A

Start better: this is one of the improvement area. Yes,.

C

Yes, well, it's it's really that, in the sense that you have to afford bare metal to be able to actually use that correctly today, now.

B

C

The kind of markets that we are thinking about, um they would not want to run nested anyway, but yeah.

B

But for the marginal.

C

Market, that's clearly a problem.

B

I I think it makes sense, because if you look at openshift, virtualization or yeah things like convert, that's also something that we would recommend to have like bare metal uh to run your your vms instead of nested vms, because you will probably have also performance differences. I guess if you are using virtualization, so for me it does not it's not. It doesn't seem like an issue, especially with you know, everything that we've done to automate uh installation and upgrades of openshift open shift environmental instances.

B

uh It's become. You know pretty easy to to manage that in a mixed way.

C

So, yes, and I think that there is implicitly in what you said, there is also a question about performance and resource utilization, etc and clearly that's a question that many people have.

C

So there are a number of important differences in how this runs and, um and so most of the performance impact actually is not due to the head of virtualization, but to the fact that we run with a different configuration to give a simple example: when you run a container today on a machine that has, for instance, 64 cpus it unless you specifically say to limit resources, etc, it potentially has access to the 64 cpus, even if it's indirectly, for instance, while it's doing ios, some of the I o completions might happen anywhere on the 64 cpus.

C

Now, when you run in a catalog containers, you're going to start a vm and the vm by default starts with two cpus and very little memory and then based on the spec that you send. We are going to hot plug additional cpus additional memory, and so on. So initially you start with, for instance, one cpu and two gigs of memory, and if you have something that requires 10, cpus and 40 gigs of memory.

C

On top of this we're going to hot plug that okay, now, let's say that you just added four cpus, then that means your workload, even if it's not restricted will never exceed for cpus and.

B

C

So, in terms of risk, it's you have a better control of resources, because now it's no longer using you know running wild, potentially, uh even if you don't put limits, but at the same time um this means that if you just have transient bursts, that could actually benefit from extra cpus. Well, they will be clamped by this number of cpus and the same with memory.

C

So that's one aspect of of performance overhead.

C

Another um aspect that is important is that the virtual machine monitor itself or the virtual machine requires resource some resources for itself. So that's why we start the pod with one cpu and two gigs of memory just to to get started, and you just can't boot linux. If you don't have a minimum set of resources.

C

Now the overhead we measure, depending on the channel, tends to be in the 300 megs range per pod. So that's not negligible for very small workloads. If you run just a basic engine x, that itself uses eight megs uh having 300 megs. On top of that, may not make much much much sense um now.

C

The problem in accounting, for that is that these 300 megs include, for instance, things like fiber for cash and other things that are really useful for your workload and again now they are not shared uh when you have a fight buffer cash on the host, you share it with everyone else. In that case, it's dedicated to you. So in a sense it it can give you back some of the performance you are losing, for instance, from uh doing the the the disco of our this, um this extravagator ufs layer.

C

So what this means is that when we do performance studies, we have performance results all over the place, and it's really something you it's a completely different configuration, so the performance results you get uh in few cases can be better uh in most cases are worse, but it really varies uh by how much, uh for instance, um you know if you do very short packet network packets, but you tons of them that that's typically not good in virtualization and that's true for any kind of virtual machine, not just in canada containers.

C

So this is really something that you probably want to measure. If you're, considering using cada containers, you should measure to be able to evaluate how big your vm should be, how many resources you need to allocate to the workload etc. So you need some tuning for that. That may be completely different from running it on a normal cluster.

B

But yeah again I think it's just a a question of when, uh um like the confidentiality, the necessity for the isolation, yes, which may be, I mean if it's mandatory, then you have no, I mean there's no question you just have to to uh correctly, as you say, um define the requirements to have a good performance.

B

But um if it's not necessary, then you probably just go with the traditional containers and.

C

B

C

One example to to illustrate what you just said: uh one of the things we are working on um that is not completely working yet because of uh some issues with how cutter containers deals with huge pages at the moment.

C

But the use case we are working on is supporting a gpdk nsio v in in these workloads, with assigned uh virtual functions from the host.

C

Now, if you use something like the siob plugin in openshift today in in theory, it will give you a given card and you're supposed to access that card and use it, but it's actually relatively complicated from a technical standpoint to guarantee that you will really restrict yourself to that car.

C

If you have the slightest container escape or something that some part that runs a previous container or whatever you could actually configure the hosts uh the way you want and essentially steal the cars from your neighbor or whatever, and so that's uh that's a bit of a risk. That's okay! If all the workloads that run on your system are yours, but if you start wanting to run some customer workloads aside alongside with yours, uh let's say you are in a telco environment near the edge or whatever there are.

C

There are cases where you want to run workloads that do not belong to you and you want to make sure that they they will not risk damaging. Your main feature, your main function of transmitting data, so, and so um so, in that case, you can essentially give a card or a virtual function to your virtual machine, run. Dpdk inside and so dpdk is user space networking and it so essentially you have no other head because your dpdk drivers are going to touch the car practically directly and so in the experiments we did.

C

We had something like less than two percent of our head uh on on some. You know on uh tests, doing uh bandwidth, tests, etc.

B

Much higher I mean, but you know that the card is only accessible from.

C

That but yeah exactly but now you know that you can't touch any other card.

B

So you're essentially.

C

Using that function, so you are saturating your 10 gig link, but it's yours. It's really only yours, and so that's the that's the good news uh of where we are going on the kind of of features that we mean behind isolation.

C

Now again, um this is uh we we had initially scheduled that srv would be there for uh for 4 11., um it's going to be late. We'll probably have to postpone that to 412 because of some unforeseen bugs we discovered in the the way huge page use. Phases are set up in killer containers today, which is it's not the right way, so you only have an insufficient number of huge pages in practice, so that doesn't work.

A

Christopher you mentioned more times that the word security I workloads isolation in order to protect um how to have an effective but basic, secure way to run multi-tenants workloads across tenants of the ship. Is there any is any new uh something that needs to provide us in terms of trends, but we are going to plan and shift cada containers uh from security perspective.

A

What mallet, when you have to run different workloads on openshift.

C

In more multitask way, so one of the things, for instance, that is difficult to do today, is if you have tenants that have specific requirements in terms of canal tuning, for example, so they want to pass specific parameters to the kernel they want to resize whatever. So so, if they want to do that today, they need to dedicate nodes to that.

B

C

A classical openshift environment and uh and that tuning may actually be detrimental for other applications and so maybe uh that's uncomfortable, and they have to essentially dedicate nodes to this particular tenant and they can't run any other workload on it.

C

uh So this is the kind of configuration where catalog containers can help, because now you essentially as assign this channel parameters to this past specific workload and only the parts with that workload are going to receive these parameters.

C

This is something that has explicit support in catacontainers, in terms of, for instance, you can specify annotations that are dedicated to color containers and that will change the boot parameters that you have all the I don't know the options that you pass to the vertex ufs demon, all these kind of things, so you have multiple tweaks. You can use along the way like this to really.

C

Configure the camera the way you want, if you want to run a kernel with traces for debugging purpose, you can do that too, and that's something that you obviously would not want to do on a production system uh if you had to do it on the host. But if you just do it in your vm, that's fine.

B

Okay, so you mean you, don't you don't have to to to make the the parameters change on the node itself? You can do it inside the vm.

C

B

Be ephemeral and then you can go back to yeah.

C

And uh and that's only this vm and it doesn't impact anybody else's workload, including other vms, so you're free to really you know, that's the same same thing, for instance, with ebpf ebpf is something that typically is restricted on normal nodes, because there is too much chances of getting it wrong or degrading the performance of others.

C

But you can safely uh do any kind of send any kind of ebpf program in your vm. It doesn't impact anyone else, so so you're free to get that capability and use it.

B

So is this all already available or is it a future implementation.

C

C

That works better upstream at the moment. There are some last time I checked, which was two months ago.

C

There were some issues with user space versus kernel in our course, and so the end result was that the capabilities were there but could not be named so it was, you could not actually specify them in the so there was a bug there. uh I've not checked recently for that, but for upstream that works.

B

Something about confidentiality, that's the part of the question, for you is that right.

A

Yeah that I was, I was just pointing now to the uh confidential computing. Something like that. I read something: is there anything related to computational computing in the future of other containers?.

C

So let me share my screen again.

C

So um there is this project derived from cada containers called confidential containers and, as you can see, it's really active.

C

uh It became an open uh sorry, a cncf sandbox container sent off sorry a cncf sandbox project uh very recently, um so sandbox project for the cncf means the beginning of a product that has nothing to do with the sandboxing in open system, box, containers um and so confidential containers will take advantage of the features that are arriving in various cpus, like amd scv in apex, cpus or intel tdx, or on s390 there's similar features which allow you to essentially encrypt the memory of your vm, so that someone on the host dumping, the the whole vm memory, for instance, cannot see passwords or other secrets that can be uh in the in the vm memory at the moment.

C

So what this allows you now is that in a multi-tenant configuration you can have two competing tenants on the same system. You know they are not going to uh be able to read each other's data, which is important, but my most important. They don't even need to trust the cloud provider are running on, so they can essentially run on an untrusted host.

C

Now these confidential computing technologies are really complicated to set up one of the things, for instance, that you need is something called attestation which lets you validate the workload that you want to run and check that you're running what you expect to run so essentially, it's a cryptographic exchange that ensures that what you're running is exactly what you want to run, and so this attestation mechanism exists in all the platforms that allow any kind of confidential computing, but it's not implemented the same way, and one of the benefits of confidential containers is that we are trying to hide all this complexity so that, essentially you get the same benefits as for openshift sandbox containers, where you just use your existing yaml files, you just add a little line to it.

C

So in that case that would be a runtime that is confidential. You may have to add some secrets as well and essentially then your workload deploys um safely and and cannot be tampered with even by the host. So I gave a talk about this. uh Let me check so that's another talk at def conf again, so you can look it up on the web.

C

uh So if you look for this, this talk, uh there is a youtube video that uh that talks about it, and so essentially the the problem, as I said, is that the sandboxing that exists today is really. It goes one way it protects the host from containers, but it doesn't protect the containers from the host. And so, if you want to consider your host as hostile, then you need to find a way to protect from that and that's essentially what confidential computing does and it's more than just encryption of memory right.

C

It's uh it's also things like making sure you cannot corrupt the state by injecting data in it, and it's also uh this attestation service that I was telling you so that you cannot replace your valid nginx image by some alternate version or whatever yeah.

B

Yeah regarding the attestation: what is it going to check like? Does it check the yaml resources to verify that.

C

No, it's it's really. It's really measuring the uh what you're putting let me see. That's that's here so essentially your you're measuring that you know on which host you're running and that you know the stack here.

C

We can extend that to also include the container image, but for the container image we typically use something else which is uh cryptographic signatures that already exist, so we don't need more than that for the container itself, but it's really checking that the candle you run on is trusted and the hypervisor you run on has not been tampered with.

B

A

C

A

C

So so the part that we really deal with is this one. That part is supposed to be something that you can validate elsewhere. That's the secure boots and this kind of of things that exists already.

A

So and everything is totally transparent in terms of the developers perspective, so they don't need to do anything in order to exploit or leverage this feature.

C

So we are trying to make that transparent. It's not very easy and that's sort of you can see the the complexity of the process here where, um when you start your your vm, you have to, for instance, send a quote. So the destination agent is going to send a code based on measurements. It did here that code itself um has to uh be encrypted with data.

C

That includes unknowns so that it's you can't have replay attacks and so on, and then we are going to use an attestation service that will let some external relying party deliver, for instance, keys that can be used in particular to decrypt the container images.

C

So, as you can see, it's a relatively complicated process and we are uh moving step by step in this. What you are seeing here is already completely outdated, because we no longer rely on machine scope. That was really the first steps uh in the process, but we now we we use something.

C

We there is a lot we can do only with oci ocr crypts.

C

So but but the the overall process remains the same.

A

We do have a, I don't know if you have anything to add to relate to the computer, which is a great news. I I just read something about confidential, but uh I was not aware that as a redhead, we are going to work on this.

C

So what I showed you here is only upstream, as I said, it's just a sandbox project inside the cncf. So that's the very beginning of the process, but I'm talking about it here, because I want to invite anyone who is interested in these kind of topics to join us and contribute, there's quite a bit of work to do and, as you see, there's already a large team working on it upstream with companies, including ibm apple and group, alibaba, etc.

C

So many of uh of the contributors are actively using kela containers today in production and want to benefit from these new features. But I see that as a really important use case in the in the future of color containers and therefore I suspect that, probably in two years or something like that, we're going to start talking about it in the context of sandbox containers.

A

Yeah sure- and there are a lot of domain where probably confidential computing or containers would be really useful, like I'm involved in some project dealing with the cloud sovereignty, which is specific for your anemia and data, there are a lot of security levels we need to deal with and the data in use. So there is that the rest of the motion that the user is one of the era where still there is a lot of things to do so, definitely a great topic um jafar.

A

I am seeing finally a question coming from our one of our attendees, uh the the the name is, it seems to be sh eris. I don't know how to pronounce it. Sorry.

B

It starts, but uh we it's just a guess, but it's a sort of cherries and.

A

The question is: is it possible to create hardware profiles and compute nodes to advertise where they can support the certain contact container workload like 10gb gpi, physical ram things like that.

C

Yes, so so I quickly alluded to that earlier that today, the the option that we give to deploy the operator itself is, is you set a label on your nodes and you have to set that manually and you can use whatever script or whatever now the the standard way of doing that is with the node discovery operator and um and that that lets you automate that, but the integration with uh openshift sandbox containers is not complete yet and so and yeah so for gpu workloads.

C

uh So that's typically one case where you want to discover um if, if this given workload, has a gpu, this your node has a gpu and um the there.

B

C

So so there are folks running dedicated gpus in in kerala containers.

C

It's a bit tricky to do at the moment, and so we have relatively good instructions on the upstream project on how to do that, and you have to follow these instructions in order to be able to set up both the node and at the the workload itself to be able to take advantage of the gpu.

C

So this can be done, but it's uh uh it's a bit tricky part of the the the problems described in the upstream documentation are related to which channel to pick to have gpu support and that's something that doesn't apply in the case of sandbox containers.

B

C

We have only one calendar and it has the features enabled, but the rest of the description does apply.

B

C

And and just so.

B

Just it's maybe unrelated, but in case of gpu, for traditional containers, as you mentioned, there's the the node discovery operator that automatically adds annotations to the to the gpu, enabled hosts and they're. You know like operators that take advantage of that, like the nema operator and such things that allow you to automatically configure the nodes for gpu support and such things.

B

So uh I would say I don't know how this relates to the containers world.

C

So, essentially, you need to have, cannot support for uh uh vfio and uh so.

C

I'm not sure exactly how this is configured in all cases, but essentially you want to be able to assign the gpu yeah.

B

C

B

So if this is configured on the host itself, because, for example, the the the the node discovery operator, if, if it configures the node with gpu support, then that would be inherited from a category that doesn't change the kernel parameters. I I guess.

C

That's not completely sufficient because you, you really need to pass specific boot options to the channel to make sure that you can actually do what you want and part of the tricks also are that there are some steps that I'm not sure we automate today likes it's possible that the device, uh the way you present the device inside the guest doesn't use the same name. So you essentially need to rename it in the.

C

The guests and that's part of the work we do. We automated for the sio v case, where the.

C

Siov device plugin tells the workload what to use.

B

C

Giving a pci address that is not valid in the guest, so we have to essentially and that's given through an environment variable, so we essentially have to rewrite the environment to rename the pci device and replace it with the one that you have to use in the guest.

C

All right so, as far as I know, that's not automated for gpus, so you probably have to be a bit uh to be hand holding the workload there.

A

So jafar, I don't know if you have a, is there any more questions from our attendees.

C

So there's do openshift macboon configs. What is macbone.

A

Yeah microphone, probably.

B

Is it like a misspelling for machine copy or.

C

Oh machine config, maybe I.

B

Don't know yeah, maybe it's a misspelling, so is that machine do.

A

C

A

B

C

Yes, so let's assume that's my the machine config question um when we deploy the the operator, so so that's a piece that is also on the web. Let me just so one of the key components or contributions is called the sandbox containers operator. Let me share my screen again to show you.

B

C

Yeah, so that's um that is the the sandbox containers operator source code, so you can look how this works inside and I don't know where this is written exactly, but essentially this operator when you're activating it with a keller, config custom resource, as I explained at the beginning.

C

So when you do this, this operator deployment it. It will actually install software uh in the various worker nodes.

C

So we use the extension mechanism for that and essentially we're installing uh a new copy of our own copy of qmu plus some other things, so some specific rpms that get installed in your uh in your nodes and- uh and so there are reboots involved, and so the machine config operator is dealing with some of these aspects for us okay. So there is a relationship with uh the mco, but that's mostly for these aspect of things.

C

um After that, I don't think it gets involved anymore. Okay,.

A

So um I think we are up to the hour thanks a lot christoph for the effects: hey yeah, more or less three other.

C

A

Yeah yeah, it went very fast right um jafar. Do you want to add something more or.

B

No, I think it was a pretty interesting.

A

B

Thank you very much christoph for the explanations and the demonstration as well. We've.

C

Seen from the demo.

B

And although the way things are underneath are.

C

B

Different, it doesn't really make a big change from the user experience perspective. So that's pretty cool because just adding the annotation allows you to.

B

uh Yeah but other than that it was very informative. I wasn't really well-versed into kata containers yet, but I think I have a better question. I hope it's the same for our viewers as well. So thanks a lot.

A

Yeah uh just a reminder for our athletes. I please join our developer sandbox, it's a great way to start to play with openshift in a very simple way, so you don't need to create an environment.

A

Just a very small one will be provided to you and you can start to deploy your your application, um a reminder for next session. Next week we are going to have a guest from a mongodb and we are going to speak about the red dead of the ship database, arches, which you said new services. A new feature would allow to.

A

Connect to the application to uh db like mongodb one in a more uh in a similar way and again thanks a lot for your time. Thanks christoph for the session was great. I learned a lot of things mainly when you spoke about confidential contains. I will take a look at some something on google just to learn more. So thanks a lot again.

B

Yeah and please do not forget to check the openshift tv uh schedule, there is a lot of rituals as well happening.

B

But again thanks thanks christoph.

C

Thank you very much and talk to you soon.

A

See you soon bye cheers.