Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2018 (Shanghai), 21 Apr 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Deep Dive: SIG Scheduling - Da Ma, IBM

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Deep Dive: SIG Scheduling - Da Ma, IBM

Please join us for in-depth understanding of Kubernetes Scheduler and its advanced features.In this presentation we talk about the internals of Kubernetes Scheduler and how it keeps track of the cluster resources in order to make scheduling decisions. We will cover the scheduling workflow and the life of a Pod during a scheduling cycle. Then we will talk about some of the advanced scheduling features that enable users to customize placement of their workloads in Kubernetes clusters to achieve higher reliability and/or better resource utilization.This session is most useful for cluster admins or those who want to start contributing to Kubernetes scheduler.

To learn more click here: https://sched.co/FuLj

A

Yesterday, yesterday, I have made it intro for you and I've walked through the.

A

Scheduling part and for the tip type session, I'm going to talk about the default and the features and some progress in this project. For me here my name is Maura and I.

B

A

The co-leader of the sick, scheduling and I'm working in my way now to work on some resource.

B

Scheduling so here's the agenda.

A

C

A

Agenda I think you have heard these parts on the different occasions like in highway highway and other companies, so I will walk through this part quickly and if you are not familiar or it's really interesting this part after the session, we can have further discussions.

B

And I will talk about the.

A

B

A

B

A

In the cooperage, this part has been covered a little bit so I think I will elaborate further on about the count scheduling some notices.

B

A

The particular thing we need to pay attention to another thing is.

B

Portal scheduling policy.

B

A

Thing is that a scheduler.

C

B

Been introduced to me, the intro this this one is also referred to the one.

C

Counselor and discover each other's here because they are the my features, updated post-graduate other features.

B

C

Not been integrated into the sake projects woohoo, they are reversed by those who are interested in this in the company.

B

C

In a substrate in.

B

The package support.

B

C

Be ordered in the priority, but it doesn't have any order. It will be. The first come first rule, and this note- and you see that's hard for you.

B

C

Mouth with our two part- and that is kinetic and it will be run and this one is done by the kubernetes and scheduling. Well, not the work with us got a writer and it.

A

B

C

To the note and a study.

C

I what you say that is.

A

See that I scheduler.

C

Just information.

B

If they are in Barcelona, consider.

C

B

C

Attention to that.

C

C

Machine has already been shared and used our record.

B

In the scheduler information.

C

In the API.

B

C

That the account of the port and the number of the pattern.

B

C

Note and then they will calculate in this country. They know that how many posts- and this notes are idle, for example,.

A

If they are a lot, but.

C

From this year, I need to calculate lay by yourself. There's no one place to tell you that how many posts and notes are idle.

B

C

Sporty color has already been complained.

B

C

B

Time it will just scatter one Pro. How much means.

A

C

Will check that which mole is most suitable for the coat, so this study for scheduling work- and this is the kubernetes scheduler Warragul- and this kind of the individuals gotta learn is way. These are kind of the waste or to.

B

C

The coop scheduler order these capture, individual and they'll just chat with status of the notes.

B

C

First, one is the predicts and it says that much.

B

C

The post and another.

A

C

One with the cross means that they're.

A

Not available so.

C

Two three and four are useful in our vailable but which one should be used and which is the priority. It is a bun by another features.

B

C

The no.3 has the square of 90.

B

C

Pod will be mounted to this.

B

C

That my two steps, this.

B

Is how the flow goes, kill.

C

All these kind of big you are the path names of the notes. Now you can see that what are the notes are talking without any code, so without any jobs, because there are surveys considered if the no name has already been considered and in.

B

C

B

Is panting at or empty.

C

As scheduled, we'll just go through the logic, the large egg, to stop two stops and choose the one with the highest priority. You can again choose only one machine, so this is the main thing. That's.

B

The schedule what a discovery is doing the.

C

Difference is that maybe the policies.

B

For prediction.

A

Authorities are different and I just mentioned.

C

That and finally they, if they choose no 3, you who saw was 90 and they just added to the server.

B

B

C

Be changed to the path free.

C

Or the post ballot should not be integrated to the communities, so this is a design versus.

B

C

Policy of kubernetes, so when you decide your schedule, please follow this rule at least all the world of the priority. Maybe if there are some changes and we.

B

C

Be really hard another one is label. Lego is quite for me to buy all of you. For example, what kind of the.

B

Notes and look at the policies that we should choose.

C

This is what it is. We don't use start of our mouth, but very much present know where we choose the note.

C

We don't use the selector very directly because it has already been replaced by no structure and we also use the post line as well. So as well as you know that there is a function, it's.

B

C

Already familiar with these, and after that, they were just.

B

C

B

C

Don't use the note very often a different server will run it on different server, and so, let's just have a look at a 10.

B

C

The poet must be astir.

B

Building two different dishes: they.

C

Need to calculate the priority so.

A

They try the best. Do these two billion to more than one machine and I just mentioned.

C

C

Here we are, so is a and then the boats will be distributed to the note 1.

B

C

B

C

Has iron, which will result in this kind of.

B

C

If we have three machines.

A

And three notes.

C

But if we have another.

B

Note it will be shown here. Another mode is the preferred mode in affinity.

C

And he affinity, it will be sad, it is scattering or preferred.

B

But it is preferred.

C

B

C

That it may have two parts, because that they have the highest priority.

C

And then all these three are almost the same and they were definitely choose one brand. One preferred.

B

C

This is the pre first data.

B

B

C

This is not infinity.

B

Cheaper solution and.

C

This is about toleration, and this was added by the latest released and I want. To give you an example, the memory and the CPU, the memory GPU.

B

Means that price.

C

And other support and the power.

B

Must be applied.

C

To one machine, but it only used for G memory. Theoretically, it should be.

B

Don't worry in here because.

A

After this memory.

C

Has been used and nobody can use the memory here as well, so this is what I do and to do the.

B

B

C

B

Did you know what it doesn't have any memory.

C

B

A gigabyte memory, it.

A

B

Can be used ever.

C

Before so, what it does is to add a tan. It says that it has a CPU. So if you don't see P you you.

A

Shouldn't be it 22, not.

B

C

Then the power should be deployed to another machine and.

B

C

Can see that the memory.

B

He meets the requirement if.

C

It has another party with the 40k by memory and want to be you evoke a deploy to note one.

B

C

To add a toleration and you need to add the GPU and plus color screen.

B

C

That's the Toleration will be increased and just scheduled.

B

C

Currently the default Toberman, which means starts the tank, can be added. Finally,.

A

By yourself, what is w? It means that, for example, I.

B

Have my client as well as I, said that.

C

I have aggregated.

B

To operate them no and I.

C

Can add the talent in.

B

This way, I just.

C

Means that, as always,.

B

I think that national TV sees.

C

Nobody can, you said if.

B

I can I can.

B

C

As a toleration, so it cannot be used, so it would be deployed to no one and later I would like to give you some more details about which is the note. You see a.

A

C

Do with Eastasia do with this. This kind of thought, 8 and the Toleration not can be added arbitrary because it can provide somebody to occupy is some machines and then others of the machines will be idle. So we need a manager to.

B

C

B

All three machines and it fraeno to.

C

Talk about the priority and preemption, and this is the one who are you, what by television it has already been available in the 1.10 version, but currently parody released. The.

C

B

Poder may be a very important one, but.

C

Oh, the machine has already been.

B

B

C

There is no way to schedule the important.

B

C

This one, the priority class to be a to make a joke. It's higher in a level regarding to the priority red, free, colors, the red means that hierarchy and the yellow one is the priority.

A

Of every no proudiy.

C

When there is a red one, they will add each have the notes.

B

Under three machines, one.

C

Each has it has different manners. The first one is that it wasn't as efficient visions as well and the second one don't cry the part of any ability. The PDP, because.

A

It has already been ranked it.

C

Has its own priority so there will I defy the corresponding information engineer.

B

To deploy to the corresponding position in.

C

This picture yeah well, I, just choose it from the lowest priority.

B

C

Discussion, if you choose the lowest priority.

B

C

Need to choose from the lowest one! Otherwise, if.

B

The higher priority.

C

Has already moved and they need to be reschedule again, so we need to have to deploy it and replace the lowest priority, one so beside the priority and also consider the affinity they don't want to break the affinity.

A

This is the French and authorities basic.

B

A

Here, actually, some notice here something special.

C

B

A

To first for this priority in the queue, as we know,.

B

A

Scheduling, there's pending queue and in this queue there is us it is arranged by authorities boundary. So when you check the result, you need to be aware that, because.

B

A

Server, so in the development it involves components such as a server controller and the scheduler there may be risk connection between them. In other words, when you, when your scheduler is considering the pending cool, similar tasks ever only considered current.

B

A

It will, in main that, when you run the check in may select the medium priority part first and then see.

B

The high priority poverty later it's.

A

B

A

It's not completely synchronicity synchronized.

B

A

When you have the additional preemption you need to pay attention, it can be a bug.

B

A

Because of the scheduler can only get the current snapshot and made the decision based on that.

B

Snapshot, therefore,.

A

You need to pay attention to that. When you see this, you need to check the blog or the monitor, to check the stators.

B

A

This is much about party intervention.

B

A

These two are in paid enabled in Veda, studying from 1.2 to 1.3.

B

A

Now, if it is lifted to 1.2.

B

C

A

By default, priority.

B

A

A

Prevention and a priority, these should be considered as two separate things, meaning that preemption means that when I think the Machine a gets extra resource. Sorry, when a pod, when.

B

A

If a has the higher priority than P, then I can't seize Beeson resources.

B

A

We have the pipeline, and now we are working on a priority as a pipe a a.

B

A

Now what I can take over is for them to kind of preemption for days. I said the hot code, power of friendship. Parity second, is when I use the job or use the PO group.

A

When I found that I get more resources than P, then B should be.

C

A

To see stellar associate because the the boat both should have four house resources, half half and half, but now one gets 60 the other get a 40, then the 48, the one who get 40 should be able to seize them resource resource. So priority is not the.

C

B

A

We are working on this direction or by thinking about doing this. So when you make your own scheduler, maybe you can try it all for as well.

B

A

B

A

A lot of discussion about what does it do for ground scheduling the cluster size is fixed by defaulting on the cloud you think the car is limitless. We will exclude such scenario. Even if you really run the cloud limit Li. Is there a cost issue now for a county.

B

Sheriffís schedule involved in.

A

This machine, this part, needs to run.

B

A

The tenth floor and other.

B

A

Jobs need to run together.

B

A

That the task is get home or completed, otherwise they wait for each other in the worst.

B

A scenario when.

A

You, the order of the party, is changed. There may be some changes, meaning that each some jobs only some jobs.

B

Alone and others that.

A

Means the green task is waiting for the start.

B

A

Green wimp are the two stars that was caused.

C

A

Waste, of course, you can find a way like a time out to fix it like after one minutes, they don't start.

B

A

Restart it or we run it, but still.

B

A

Trauma the same problem may occur, so we use them in hands feature to put down schedule.

B

A

B

A

Can run together when we try to work on it, we realize that we need to consider a lot of the cases. For example, here.

B

I know if you can see it here, this is.

A

Light blue these parts are lights, light blue. That means that when we run in Sparks the install sparks are three jobs. There is a master.

B

And a workers note so.

A

Not every job is the same and.

B

A

This part there.

B

May be MPI.

A

Worker and plan together and.

B

A

There may be one or two master and several workers.

B

A

Master in the worker.

B

A

B

C

Not be the same image, a.

A

Deport may be different, for example, in the t internal, but they are different here.

B

A

We are required to make them compatible to consider them otherwise, and they won't be able.

B

To run it to learn in.

A

Order to make them compatible use thing called a product loop.

B

A

This is the port group beginning everyone to call it job, but Java has the best shoe. You can find it as something.

B

Also called a job, so we call it party group.

A

And the definition is at a finalized, a stage.

B

A

Find our tolerance resource and I will talk about it later there. It says later.

B

A

It work not sure if you can see clearly how.

B

Explain all the way from.

A

Here to the appear to the bottom, to.

B

The cards before the.

A

Red box, it is a native.

C

Adventure yeah file and in the red one red boxes.

A

We add a connotation scheduling, kk8 s, dot.

C

B

A

You see here is the assistant which are format and for the bottom red box, its CRT, which is called up cords. What a group, the definition or.

B

Slower the particle as I said, there's.

A

A name as long as these two names are matched in the scheduling you would be recorded, as one part can create it. In this way you can actually create multiple jobs or jobs plus RC audio-visual. As long as your connotation of the group name is consistent, is the same. You were blown to the and in the crowd or the group, it will have a mean member.

B

A

Me, member what I said and these six thing this takes six.

B

Members and you can set our six.

A

In the scheduling that means in the scheduling when we have the six members, then.

B

It will be scheduled this.

A

Is what can our scheduling.

B

C

A

C

A

C

A

You can see a mean member is eight and on the rights main memory, six. What is really for MPI task.

B

A

Task flow change.

B

A

B

Model, meaning.

B

A

Pending thought, the.

B

A

Pod to be run, but.

B

Not future, but for this box.

A

The extra team.

C

Would be nice to have meanings.

A

B

A

Far, as can run the 5s cuter and the remaining too nice.

C

B

Can run in order.

A

B

Support both model here.

A

What do we do? How can we do with the remaining two, because.

B

A

You have multiple tasks, and these two would work on sharing use the.

B

A

Percentage and the previous six one six months to meet to maintain.

B

A

Basically comes give account scheduling roses room in member, which is quite simple.

B

A

Long as you get the enough resource.

B

Then you can run it. What.

A

Is it before? Is things done in gas gauge scheduling.

C

What one problem is.

A

That you can, then you can do.

C

B

A

B

A

B

A

Tricky thing is that the quarter.

B

A

Quarter size can meet the requirement. You are not able to create extra pending scheduler.

A

We can't do anything.

B

Nothing because if you set remember.

A

As 8 and the above as 6, you can schedule.

B

All to manage the whole region here for.

A

This part, because the photo size is full and it won't create spending pods.

B

A

Are several ways.

B

A

Beginning you can set a figure quarter, then you will fix the problem.

B

A

Accommodate all the bodies.

B

A

Can just set the quota and your settle at Quora big enough the world, but once problem is, if there is too many parts, it will affect his EPM server and the.

B

A

Also, when it creates to know many.

B

Things hallelujah, the quarter would be minimis.

A

But we still want to add a quarter, so we have some discussions and we.

B

A

B

A

The total resource I mentioned we will make an enhancement.

B

A

As a reserve, some of we mention about the protocol.

B

When a group is total resource and.

A

The resource will be reserved, so when the next time you have.

B

Already next time you create the outer automation.

A

B

Jeanne that.

A

Means we make a reservation.

B

A

We can avoid the.

B

A

Will wait all the resources are.

B

In place here, we will create an idly so.

A

This is what we need.

B

A

For the enhancement later in the future, so far we haven't realized.

B

This total resource, yet another.

A

Thing is about life cycle.

B

A

Easier to do because when you kill it URI studies, there is no connection between the new one.

B

A

Tricky thing here is for a training java, api.

B

A

When you cure it, you have to restart the whole task.

B

A

The assisting controller- there is no feature like this, for example, for the RTE, as here you will restart an individual port instead of the whole thing. So there.

B

Is no feature to support this yetis. Therefore,.

A

In the future running, we were at all.

B

Events, policy and.

A

A port group controller.

A

C

C

C

These widest people scheduler and we will discuss it and we have a quest la première sa, which is the scheduler framework we made based on that, and we may do that.

B

C

We hope, as we have all.

B

The resources that.

C

We can use at.

B

C

But we need to have those restriction.

B

A

B

Are namespace go into the content.

C

C

B

Also have other natural fire restriction, the.

C

B

C

Port or divinity or in Tiffany's voice on some.

B

Particular namespace.

C

And policy we can have more policies.

B

C

B

C

Out and it will work and.

B

C

Have this latter's if they manage.

B

C

See that whether it will pass it or we have to reject.

B

C

B

Idea is to control to.

C

Make sure that the policy should be controlled by our criteria. We shouldn't be more or so says from some external hats, and this is what I.

C

Think that's you know controlling with that. So I would like to skip that except I have some question marks here, because the iyammer power has not.

B

Been finalized.

C

If you will lie to you, set.

B

C

Megawatt ACE scheduler Percy will be live out of three other. It accommodates the signal all right or a dock.

B

C

We use a sturdy base approach.

B

C

Still under discussion, because it is a really important project and the policy working girl is still cooperates with other projects Microsoft and see that whether we can cooperate with order. So a lot of things are any discussion. I hope that you can change and follow up the latest data of this project and if.

B

Something has already committed and we all terminated it.

C

It's not by the policy and the.

A

C

And as a scheduler will decide whether they will have to use these basic features to restrict, attain the finishes.

B

C

The teat or something there will be those striction on this.

B

C

B

A

C

Labash and not.

B

C

Well, the Cooper patch in this part, because it will be.

B

C

To infection and then create a new one and every scheduled.

B

C

All these are some cases, some teacher offenders, because they had already running for a long time, so the status of a nose, oh, the pose, are not meeting the default state are.

B

The schedule and all these artists have a policy.

B

C

There are a lot of the Popes on this node and.

B

C

Will I to rebounce it.

C

C

For the demon, sad.

A

C

To be recreated I, never imagine if it violates that its PDP, because.

C

All the members of the rigid group of admirers.

B

Schedulers you.

C

Can scan a capo and join and this one from the PI 2 and this one from Ali and also in this work, and we also have a group from the twosome. So if you are interested and you can join school and talk about the developing of the schedule with us and so much from me,.

B

C

We still have any time for a Q&A, oh yeah,.

B

C

Already used all the time in slot, you should have so, but if you need help in session, you can have it as well. Okay, in that case, you know, let's call it today and if you have any questions and feel free to contact me. Thank you.