oVirt applying Nova scheduler concepts for data center virtualization

Video in TIB AV-Portal: oVirt applying Nova scheduler concepts for data center virtualization

Formal Metadata

oVirt applying Nova scheduler concepts for data center virtualization
Title of Series
CC Attribution 2.0 Belgium:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Release Date

Content Metadata

Subject Area
For several years now, the oVirt project has been leveraging KVM and relevant technologies (ksm, etc) in data center virtualizations. Being a mature and feature reach, oVirt takes another step forward with introducing a Pluggable Scheduling API. This presentation will review recent oVirt improvements in the areas of VM scheduling. The first part will discuss the architecture of the new scheduler. In the second part we will show samples of VM scheduling plug-ins, and integrate it to a live setup
Scheduling (computing) State of matter Multiplication sign Data storage device Electronic mailing list Sampling (statistics) Bit Number Connected space Human migration Data management Computer animation Meeting/Interview Video game Configuration space Computing platform
Filter <Stochastik> Distribution (mathematics) Algorithm Scheduling (computing) Graph (mathematics) Multiplication sign Weight Structural load Set (mathematics) Number Befehlsprozessor Computer animation Logic
Slide rule Category of being Scheduling (computing) Computer animation Weight Sampling (statistics) Set (mathematics) Normal (geometry)
Human migration Domain name Scheduling (computing) Process (computing) Weight Data center Bit Virtualization Iteration Graph coloring Social class
Domain name Filter <Stochastik> Human migration Scheduling (computing) Service (economics) Computer animation Structural load Weight Mereology Extension (kinesiology) Social class
Filter <Stochastik> Category of being Computer animation Logic Weight Structural load Chain Keyboard shortcut Set (mathematics) Endliche Modelltheorie Table (information) Resultant
Filter <Stochastik> Scheduling (computing) Computer animation Validity (statistics) Bit Field (computer science) Social class
Filter <Stochastik> Server (computing) Validity (statistics) Weight Multiplication sign Structural load Set (mathematics) Field (computer science) Electronic signature Category of being Message passing Befehlsprozessor Computer animation Order (biology) Endliche Modelltheorie Physical system
Filter <Stochastik> Trail Server (computing) Scheduling (computing) Divisor State of matter Multiplication sign 1 (number) Field (computer science) Number Frequency Roundness (object) Theorem Endliche Modelltheorie Mathematical optimization Computer architecture Stability theory Distribution (mathematics) Graph (mathematics) Weight Structural load Chemical equation Sampling (statistics) System call Connected space Film editing Process (computing) Befehlsprozessor Computer animation Logic Personal digital assistant Lastteilung
Word Scheduling (computing) Weight Lastteilung Variable (mathematics) Power (physics) Social class
Computer animation Configuration space Social class
Filter <Stochastik> Computer animation Weight Multiplication sign Student's t-test Mathematical optimization Power (physics) Physical system
Computer animation Social class
Filter <Stochastik> Default (computer science) Server (computing) Scheduling (computing) Service (economics) Java applet Chemical equation Weight Staff (military) Directory service System call Formal language Roundness (object) Computer animation Logic Right angle Endliche Modelltheorie Physical system
Revision control Scheduling (computing) Computer animation Execution unit Quicksort Endliche Modelltheorie Logic gate Writing Pole (complex analysis) Sinc function Number
Scheduling (computing) Dialect Information Structural load Physical law Data storage device Electronic mailing list Student's t-test Mass Word Data management Befehlsprozessor Computer animation Uniformer Raum Personal digital assistant Semiconductor memory Endliche Modelltheorie Physical system
Computer animation Bit
Computer animation
Computer animation Food energy
Computer animation
Computer animation
Computer animation
Computer animation Lecture/Conference
Computer animation Set (mathematics)
Computer animation
Computer animation
Computer animation
Computer animation
Computer animation
Computer animation Multiplication sign
the of the the thank you if you and so on and so on and so that you know could the we we we
we we the state of the union being used in OK you wanna start hi my name is Yolanda work for that I'm in the colleges and I'm working for the US-led scheduling thing then antidote about scheduling with a number of that ball and say basically we will talk about the
problem of scheduling a little bit then will go ahead and talk about a novel a future scheduled concepts and then will be diving into or its cage in the 11th samples the I think the best way to understand that our through the samples OK can you raise your hand if you heard about lower than against a lot of your life to for a so over for those of you are no no don't those of you don't know what is over or it is there and management platform for for all of the ends based on the Indian Life revises can handle thousands of the hands have blood snapshot storage migration alive the and migration like everything is like physically and sorry we support also advanced that configuration for those those and a lot of storage connections sound NFS glass the basically let's see what we've got a long time ago and got the questions from the user's list I urge you to use at least if you have any questions about over it's quite know quite active the how can I define a maximum number of running the endospore holes and it's for the
trivial but we've been supported that yet then so we get
back to it later just wanted to say that you should use the use of this OK what we have in our graph that long time ago basically we had to Distribution Algorithms for uh running my the ends when we selected the and
when you ran at the and then we selected a host based on CPU load policies either even distribution to be a low-pass saving and then it's pretty you know a limited we have only 2 Distribution Algorithms and we can construct a user-defined 1 basically that is they then what they the which they had you take a low with looked at a number are scheduler is we which brought us filters and weights very so an easy way to filter out there to schedule VM on holes basically the federal very cops and clear logic that gets a set
gets a sets of balls on the left hand side then we ran the feaLDA on on them and then we applied weight weights on top of them and another called
slide from normal scheduler the we collect we collect a sets of weights then we sum them up and then we ordered the holes that we got
let's see as soon as simple sample for novel scheduler and this is around filter written in Python basically it's a very simple method that gets a single whole stayed out and then and a set of properties and then run an really
simple color that I know
the decide if this this this house is in the scheduling posses or x excluding excluding the from the scanning process uh according to they request a drum for that the end OK the when we looked at and took it to the data center this virtualization which is over we saw that we have a little bit of a problem with that concept because in and an older back each filter and weight is applied on a single holes and we have a larger concept of class 3 you know which called migration domain so in that amendment from iteration domain each
be on holes can be migrated to a separate the also not migration domain that cluster and we have another concept of load balancing 84 . cluster and also a policy container which the user can decide and creates on policy and applied on the so let's take a look what we had in over basically we can create internal and external filter and weights the internal parts is 0 is and for better performance we it within the service so we get a quick access to the DB and in originally called of all the filtering weights from the previous forward Back then when when we didn't have centers in way well migrated into that internal scheduler and the extent of the you basically all the users can support and extend if we apply all the filters that it would like the future and and weights on all the holes in the cluster for better performance and we want to have a better grasp of how the company in the cluster behaves
we have a container is called class
policies for you for each cluster we can define the set of filters set of weights and a single load balancing for each cluster and then with support custom properties that you can and it's gonna pass through through to uh internally this is and that some of the new model on left on the left hand side you can see a set of posts within the cluster and then we applied each filter chain each filter 1 on top of the other then we construct the way table which gets us the result of the selected holes that month is scheduled to be a mom the a we have
the same concept that the uh is no bind filters the existing logic that we have previously
which world basically validation was and migrated into their
internal filters and we can extend it using extend you in Python using the external scheduler that I will get into later 1 tissue show you're really easy sample of
how we can use field there basically this is a bit filtering Python the name of the class will be the name the René that day and then
the server gets the the properties validation of basically a set of property that you can add within the server and then the filter we get it this is the name of the filter and this is actually the signature that and you need to override in order to extend of In order to extend a filter to add a filter to the system this is how you get this is how you get there custom properties from the heat within the filter basically until when the there is all about here you can see we get the time and if the time is within the interval that we get from the of the user then we friend there really with Brenda holes returned all the that we got if not we just every will all the holes from the field that's kind of an example to let's talk about the weights it it waits all the fields that all the holes that the pass through the filters the we had a pretty find weights that the 1st to our are CPU load filters and then in 3 . 4 the we added a
lot more weight models that's the way
the models which is kind of easy now to add because we have a new architecture each filter can have hand have factors to begin for Eric either through the solar filters each weight can have factors and their character is the fact that the weighting factors we can also x and extend weights let's see another
sample in this sample we use a connection to our to at the server using in Python as the case that we have this as the k is backward-compatible stable we connect to as decay and then the logic of of the way it is basically we iterate over all the the all the holes and we append to elicit toppled the whole state the and the the weight of the of that holes here it's cut but the holes the weight of the holes is the number of active in the ends on that also if we have a wholesome basically what we ordered by and number of the number of running the and the thing that tells and the let's continue have about In the load balancing that the that's the 3rd model we have the in the cluster policy each graph suppose you can have only 1 and load balancing logic the basically you can do whatever you want within the load-balancing connected decay and and shut down everything basically that what was appointed there is that while the lasting logic will return at the end if track of holes I was showing in the sample later on In the end we will that the and a single and and within server basically more are safe to to migrate a single and within an all as a period of time not to cause immigration last in the in use all our resources for that pretty not safe to do that the we also have a pretty fine load balancing they within over the 2 a legacy ones on CPU and now and in 3 . 4 even distribution 1 that we even the and distribution that we didn't have the same there balancing sample that I wanna show as i showed earlier is we get the same number that we want to shut down all the ends but in this example we will actually do that and not exclude also and prevent users from mining holes aviones like in the field the example is show all in that all in the wake up call for we will basically they get all the holes and if its wake up our will activate all the host if we need to sleep and there then we will connect using our as decay and get hold of the ends from that tells we will stop all the theorems in deactivated holes same then the example this is how we use it internally to migrate to the ends the you get a according to some logic you get the overloaded holes it's a so called me but then you select here it's random in the prose the and that tells then we actually printed because viewing the reusing city I'll do to get the data uh from the from the model and the return it and we're gonna right listed holes which is kind of a filter the 1st which filtering we do before we activates the filtering weights and the normal process that we do basically as I said we have a cluster policy which is a container for all the filters and weights in a single of about a balancing you logic and we have 2 optimizations for a cluster policy a speed and overbooking basically we're uh each time schedule at the end we run it 1 by 1 because we want to prevent overbooking in the same round it we want to guarantee the same resources for free and if you will and try to schedule to the hands together we can fail because the the book they both see the same thing resources so basically this the optimization is to exclude and
weighting of the variables
so later on the load balancing will be doing for us and balance out the the class the class and overbooking is to me its the war island the just be able to power lies with the scheduling in Brussels let's see if we can have show those pictures within a word I will carry on
then this is lowered the VM is you know because of Wi-Fi the piano and the the M is unknown it's running somewhere and here I
consider here I can configure new class of policy I can give it a
name like the shut down 1 I can add they external filter that I added to the system the shutdown holes filter do they enabled filters here is a weight I I will give the optimal for power saving that's try to other data although the and for within a single almost as much as possible then I will we select did that and advanced analyze being created earlier I can give it wake up Pollock 80 in and the shut down our the dates press OK should be created so the the criminal 2nd what
OK let's go back to that you know no children take some time but we have not connected to the the and as to what they can only be something happened to the began doesn't like me yeah I have students I have features that you know need to work so can OK let's think about this in strange each believe me In this work the to who were
here the only thing that the no
sphere and that the
mind was showed that and then you
go to cluster I feel like
in I think we have about to say they I
create a class of policies that have showed you then I created to a cluster that already find so all the holes within that laughter without going to that of policy OK let's talk about
how it's implemented there in no back it's disabled by default say whether wants to expand to add filters should be able to know how to install external scheduler that scheduler is a separate process is written in Python we externalize it because you want to guarantee the engine safety nets you know the user rights according to Mead the dangerous the system we want to allow other non languages well what if you know end all written is written in Java sold and this model is it written in Python and going wants to uh support staff which is kind scheduling the service because all way the it's a separate copy and you need to In many uh how it works basically it's initialize it rounds it reads from now on local directory all the filters weights and balancing logic that you wrote then it's about publishing down IPI the engine the server reads it and then it waits for the calls from the engine for filtering Wang balancing
this is how it looks like when it's loaded the filter and lower bounds in
gate a back to the users this now the can easily write filter that in button maximize the number of running the unions 1st poles the In early OK to sum it up uh we support the easy Python means for you to use it for you and scheduling uh you can I can manage a separate policy for each cluster for each group of and every version that comes out forward gets new models for skater questions this sort of question that I have to be to read what I what what I really related to this problem since you have by expressing a unit a good
right now you may have to go in
this work we have to sort through to the students with the word which stores will reach the region of the other possibilities as they were you should just work on work so all this work guys find out what the what we are doing so it's yeah and this basically because the government because you may be out of school so the regions of all through stuff isn't what's that that going to the store and think about whatever you know it you are much harder question the and I think you mean with any you know when you stand to things you can do whatever you want when you extend you know a filter or when you basically can connect you have to provide their own use any as the case like so but this provides with the the engine is all that you have a lot of memory and and CPU load it if you want to connect to other systemic provided geologists there and you can use you the we the some place and so we have a few days in the also talks open but the law and the will asking us to connect the schedule to would be C system that is monotone and it's far uniform tools for example they can monitor all of the supratemporal tool and defend speed and they can actually predict the thief the sense to its friends speed is constant all z all and the supratemporal so is high the course is going to college and done in a few minutes so what they ask for masses to give them a list of all simply connected lately some of them because they have a little more information than models to the have and there are also many other examples which of those similar of for example Cisco's Villasenor concepts the actually want to lick least some of the whole fills the middle who's going to go down to the Villasenor also metals in very brilliant oppose this is actually how you want the the value of the public to the details of what 1st you say well all the all that you know very well maybe of what was we worked the next thing is if I have this storage you information or information that had to do with the quality which storage perhaps I could utilize are moved to during showing in a storage for our storage most which was 1 of the other but could be management fees and long as it and and so on and the 2nd 1
and the more questions at the and close friend of the thank you for your thank
you for the and that a
bit and a thank you for the good news is that the the the the the
the the the the the the the the the the the the the the the the the the the the the the the the the the the the the
the the the the the the the the the the the uh the reason was that the the the the the the the the the the the a
thank you thank you and the other thank this thank
you this is the thank you same here in the thank the thank this is a this is
a this is the this and the other thank you and the other thank the thank
the if this is the mother of the the the the the the the the the the the energy the the the
the the the Chang
the other and the other thing that a stays of the same set SFS then the the the the the the the the
the uh the the are the
and thank
you you thank
you thank
you thank you thank you thank you thank you thank you thank
you and
you thank you
you thank you thank you thank you thank you thank you very the the the the the the the the the the the the the the so there the the little things along the time