Cloud Native Computing Foundation Kubernetes Community Days (KCD) Chennai 2022, 30 Jul 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Understanding Kube-Scheduler-Simulator by Pravar Agrawal

Description

Kube-scheduler is the brains behind managing node assignment to pods in a K8s cluster. And inside a Kubernetes cluster there are lot of things happening in the form of various api-calls at a given instance of time which makes it hard to visualize how pods are actually assigned to nodes. But, what if I tell you that there's a way to visualize and actually experience how the pods are getting assigned to a node and other resources like Nodes, Persistent volumes, Storage Classes are getting created with the help of a simulator. Cool right? Kube-scheduler-simulator is a new project under kubernetes-sigs scheduling which is also a web based simulator. On the simulator, you can create/edit/delete resources like Pods, Nodes, Persistent Volumes, Storage Classes to simulate a K8s cluster. It works as a web application which can either be deployed with Docker or can be run locally on supported machine.

A

Hi, my name is pravar agrawal and I welcome you all today, the kubernetes community days chennai, I'm thankful and very excited to be a part of this event, which is happening for the first time. I believe you all must be familiar with cube scheduler and I've got an interesting and fairly newer tool called cube, scheduler simulator to present amongst you before we move forward a little bit about me.

A

I work as a devops engineer at ibm isl for the ibm kubernetes service and those are my different handles to reach out to me on twitter, github and slack.

A

For today's agenda, we have what is good: scheduler simulator, its different components and its architecture, how it works and a little demo to show how to get started.

A

Kubernetes scheduler is one of the main master components responsible for shuttling pods onto the worker nodes. The scheduler may look as one, but it's made of small different parts doing clockwork to stay everything put for the cluster admins. The scheduler may always look like a single component running either as a pod or a surface on master nodes, but when it comes to visualizing how scheduler actually works, it's pretty hard to do so.

A

Well now we have cube scheduler simulator, which can easily help us in visualizing how to create edit delete of pods nodes volumes, which works by simulating inside a cluster cube. Scheduler simulator is a kubernetes segs project which is maintained under scheduling it's an open source project which is written in go and it can run very well locally, using docker or the docker compose.

A

It mainly focuses on visualizing resource creation. From kts scheduler's perspective.

A

We can simulate the creation of different resources like nodes, fonts, storage class and volumes.

A

The node assignment of pods by scheduler is done on the basis of scores and filter plugins, but in the scheduler simulator we have custom, filter or score plugins which are used in the place of default plugins.

A

We can also configure custom scheduler configuration by either passing the configuration file or we can use the default scheduler scheduler simulator has a web ui which can be used to try out the behavior of scheduler and check the working of plugins as well.

A

Using the web interface, you can create a new resource like node part, priority class or the storage class, and we can see the result of the scheduling in this section.

A

We will see the ui in detail once you move forward.

A

In today's presentation, we'll mainly be taking a look at how scheduler simulator works, how to get started with it by installing in local.

A

Let's take a look at the different components involved in the cube scheduler simulator.

A

Once the cube scheduler simulator is bootstrapped using the docker compose, it basically creates three containers. One is the hcd server another one is the simulator server or the actual backend server and the other one is the front-end server.

A

The back-end server, which is also called the simulator server, executes the cube api server at first. Since the cube api server is the brain of any kubernetes customer.

A

It is followed by the position volume controller, which will help in maintaining the persistent volumes across the cluster.

A

Then it also triggers the scheduler, which is the default scheduler, which comes with the cube shadow simulator, and then this simulator server is initiated which will help us in actually visualizing how the different resources are scheduled inside the kubernetes cluster.

A

We also have a web ui, just like we saw in the previous slide in the lcd server, which acts as the default data store for the the whole project.

A

And this is the architecture diagram of the scheduler simulator. The main components are the simulator server. So as soon as we execute the docker compose up, it basically executes the api server at first.

A

The api server then triggers the persistent volume controller, which is followed by the scheduler, which is the default shadow inside the cluster and then the simulator server.

A

These are the different api endpoints, which are also registered multi, simulator server or the respective component has been created or executed.

A

We can directly get the metadata all the other information by querying these endpoints running inside the scheduler simulator.

A

For the installation, we can simply give the following command by after cloning. The project into a local which we'll just see in the demo, which is the make docker build end up, and it will start building all the three containers which involve the sap scheduler simulator and the front end.

A

And now we'll see the demo, we will first see the installation of the uh scheduler simulator in my local and then I will run the docker compose in my local machine so that we can execute the scheduler simulator in the in the real time.

A

So I'll follow the guide, as mentioned in the scheduler simulator github wrapper, so we basically need to clone the project and then we'll simply give.

A

Command, as mentioned here.

A

Either make docker will end up or make start to make docker build and up would basically trigger the uh docker compose file and we'll try to create all the uh containers, as mentioned in it.

A

So we will try to create the simulator server the simulator front end and the sad.

A

In the interest of time, I already have the project cloned in my local and I will try to run the docker compose command.

A

It's been wait, while the containers are being bootstrapped.

A

And it is done.

A

So uh you can see that three new containers have been added just now, whereas the first one is mentioned h of xct running at the verizon 3.4, then the simulator server, which is the backend server for our project and the simulator front end where which will be running the simulator ui, so ui would be accessible and we port 3000.

A

So what I would do is, I would just reset it.

A

And I will delete the existing resources and then we will try to create fresh resources from scratch.

A

So you can see on this ui. We can basically create either a new storage class, new priority class, new persistent volume claim or the new node or new pod. So we'll start by creating new nodes.

A

So once you click on the new node option, it will basically give you an editor where you can put in the details of the memory specification or the cpu. You want your node to be configured with, so I will not change it and will try to schedule it with the default values.

A

And if you click on the node, you can basically see all the relative information with respect to the newly created node.

A

So it says that the phase is running for the node.

A

Now I will try to create another new node with the same configuration.

A

Then node 2 will also be having the same status.

A

A

Now I will create a new pod, so by default the part which will be created will be based on the pause container image, and I will just try to minimize it to some 2m of cpu and one or maybe 2 gigs off.

A

And you can see all the other default uh information mentioned, which is the uh termination message, path, uh image poll and then the dns policy.

A

And then I will apply.

A

You can see that the part has been scheduled on the node one and now I will create another part with a little change in the configuration and keep it as five and.

A

And five gigs of ram.

A

And it will also be having the same configuration apart from the resources as compared to the other part.

A

Now we'll go and see the first part which is being created just now, and you can see that the few fields have been added. Apart from the resource definition page.

A

So one is the filter.

A

By default, the filters defined for our both of nodes uh are like azure disk limits, uh inter part affinity, node affinity.

A

So these are all the uh by default filters which are defined as the as part of the cube, scheduler basic configuration.

A

And then, if we check this score so uh here it says that the uh scores are defined on few or the certain factors uh which is the image locality, uh interpod affinity, node affinity, node resources, balance, the location, resources, threat, portable spread and attain condition.

A

But in our case, since we did not define any of the uh inter power affinity or node affinity or even part of policy spread, so we could see that the values are termed as zeros in that case, whereas for the node resource balance, the location and node resources aspect, we have a default value of 96, which is a score and then at the end we basically have a sum of the finalized or the normalized and the applied plug-in scores.

A

So here we can see again that balance, allocation and resource fit are again is 96, but the power topology spread contain toleration have been termed as 100 and 200 for each node.

A

So to understand, plug-in or filter score locations, uh which we saw just now in the cube scheduler emulator ui. For each and every part, we need to understand the motivation behind scheduling of pods onto the different nodes.

A

The scheduling of parts onto the nodes are managed in two phases. One is the shuttling cycle and the other one is the binding cycle, the scheduling cycle. It selects a node for the part and binding cycle.

A

It applies the decision the to the cluster, and we also have few extension points like pre-filter, filter, post filter, pre-score score and the normalized code.

A

All these plugins are used for various reasons like filtering out the nodes that cannot run the part, some for pre-scoring work, which generates a state for plugins to use encirclement used to rank the nodes that have passed the filtering phases, the scheduler.

A

It eventually calls each scoring plugin before it binds the pods onto the respective nodes and the allocation is done based on the best score value calculated.

A

So before we move on to the next part of our presentation, I'll try to bring down the existing setup, which I have in my local.

A

And to do that, I will basically call a docker down command from the mac file.

A

And, as you can see, it has started stopping all the three containers, which is the hdd, the front-end server and the back-end server.

A

And it has been cleaned up now.

A

So, moving on to the next part, uh we have a few other implementations being done for the scheduler simulator. uh Recently, the project was earlier using github actions to run its ci checks, which were mostly being triggered as part of the pr and push based uh checks, and uh there were some works done to migrate to pro, which is a kubernetes based or the uh coupon. It is community maintenance tool uh to have the cia streamlined with the very common tool throughout the org.

A

So currently the cdi part of a scheduler simulator is being handled using kubernetes brow and the end-to-end test for the scheduler emulator, which is currently in progress.

A

These will focus in testing all the uh different uh components of the channel simulator like the cube api server or the response returned from the q api server cube scheduler uh whether there was a node or a pod created properly or not.

A

So if you are interested in knowing more about the scheduler simulator, you can always follow the project link.

A

There is a whole a lot of documentation available, which will be helpful for the beginners to get started with. There is a guide mentioned in the root of mention in the need me of the project which will help in getting started and how to use shadow simulator how to bootstrap it locally.

A

We also have a dock section, which basically explains how the scheduler simulator works, the different configuration you can use for cube api server and if you are interested in checking the responses of the different apis exposed, you can also check this section.

A

Few links to follow the project link is there. The documentation link is also there. You can always reach out the maintainers of shadowless simulator on 6 scheduling channel over the kubernetes snack.

A

And that's all for the presentation. I hope uh you would have enjoyed uh knowing a few things about scheduler simulator, it's a really good tool. If you're looking for some help in uh understanding and visualizing how scheduler works and the different components around it works, the community is always looking for new contributors who can come and pick up issues and start working on it. It will be very great to have different folks working on it.

A

Once again, I would like to extend my thanks for giving me a chance to present.

B

Hey thanks for the insightful session, it was really helpful and um in the end you said like you are welcoming contributors. So one question that I have around is: what is the skill set that you're looking for, as in like say, if I have to be a contributor, what is the minimum requirement that you would expect for us to contribute in your space.

C

So to contribute in any project under kubernetes, uh you can basically start afresh. Even if you don't know or go that much or you're, not that much aware about kubernetes, you can start helping with the documentation or maybe bash scripts which are used for end-to-end testing or cia cicd path, so there is always always way up the ladder from there, and if you are really interested in uh go development, you can pick up few issues which are around uh for helping with the goport, otherwise in any of the project under kubernetes.

C

You can find your way in for scheduler simulator. uh Similarly, you can always help in the documentation part. There is front end part uh which you must have seen just now in the presentation. If you are good with javascript or if you have keen interest in that, you can also pick up issues around there. Then we have apis written in go or you can pick up issues around there as well, and there is a lot of work going on with respect to end-to-end testing, uh which I mentioned.

C

uh If you really like to have, you know just knowledge around trash scripts, you can pick your way up from there as well.

B

Sure that answers my question thanks a lot, so participants like do you have any questions. Bravo is around and he's dating he'll be more than happy to help. You guys.

C

So uh one thing I would mention is, uh uh please take part uh in any of the uh community meetings uh for six scheduling uh those happen uh every other thursday around 10 30 pm uh india time, and there is one which happens uh for the apac or the asia pacific every month, uh the first thursday of every month. So you can reach out to me on slack. I can share all the details and any other community member which is available on slack.

C

You can always reach out for help and please don't hesitate to ask any questions. uh The community is very friendly, as you must have been aware of.

B

C

One of the participants.

B

Is already like he is interested to contribute, so uh so you've already answered that you're, reachable in slack. So is there any other channel that we can reach reach out to you or is it is uh only slack or you are you're active in linkedin.

C

Yeah, I'm actually on linkedin, uh my gmail email address would be there or I can have them via you. The other ways would be twitter. I'm active on twitter as well these days so yeah cool, I'm like all these.

B

Three four things: I have a personal question, I'll tap it up and have the last question. So I like this idea of simulator. So how what was the starting point as in like what was the spark to come up with a simulator because generally people don't we, we all tend to work in cli and we generally don't, or rather not have the requirement to have a simulator as such. But how do you think that you want to come up with a simulator and where did this idea come up from.

C

So uh you know, I may not be the best person to answer, because I am not the one who started this project, but it basically came out as a gsoc project and the main idea was to have a showcase kind of uh which can help people to really visualize uh how the different plugins and this different scoring happens when you are trying to assign pods to a different, node or like or to distribute to different nodes uh from the scheduler side, because we would always read in the documentation- and we can see in the cluster that the part goes to node a node b, but in actually how does it look like if it is to happen somehow right yeah?

C

So that was the main intention behind it. uh The project is still growing. uh We are adding different new components, uh different uh support for different custom plugins. You may be able to write your own simulator one day so uh yeah. It is growing from there sure.

B

Sure, thanks thanks a lot private, you have answered all the questions and the session was very insightful and helpful. I hope.

C

Your participants.

B

Are benefited as well. Thanks thanks a lot. Thank.

C

You have a good.

B

Day, bye, yeah.

C

I'm you too enjoyed yesterday yeah thanks.