Internet Engineering Task Force Interim Meetings, 9 Feb 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: COINRG Interim Meeting, 2021-02-09

Description

COINRG Interim Meeting, 2021-02-09

A

So now we're recording, and so we can go to the agenda um so we're.

B

A

uh Four uh presentations and mine will be shorter, shorter. I don't know how I'm going to deal with the nsf thing, but anyway, uh we've been very lucky that we were um asked by rakuten the the cloud operator um to maybe collaborate to their initiative on autonomous networks that they want to start at itu.

A

Because of that, I asked leon wong who's, going to lead that uh effort to present this morning or whatever time the day it.

B

A

Then we have professor alex gallis, whom a lot of you know from his, uh I would say, stellar um experience and um you know everything he did and he's going to review.

A

uh 40 years of network programmability- and I myself am looking forward to this because I met alex last year when I was doing a keynote in a conference. No another keynote. I was on a panel in a conference and alex had a lot of very good uh background on a ton of stuff that was done in the past in uh network programmability and how maybe what is happening right now is going to be different from the outcomes of the previous uh attempts.

A

At doing that, uh then uh we're going to have a presentation from edgar ramos from ericsson, uh it's a paper that we did together on descriptor for intelligence services and, as you know in this group, there's been a lot of discussion on the use of discovery and data discovery to allow network programmability and allow intelligent services inside networks. So I'm sure it's going to be. um You know it's probably going to give rise to a lot of uh questions.

A

uh Then we have a paper from radistan in cambridge. So I think we have a uk dominated meeting today uh on montana programmable switches, which is work that he's done with noah zilberman, who have a lot of you know, because she is a participant in this group and she's been very supportive.

A

And then um I don't have a lot of slides. But I will talk to you about a workshop that henry schultz in a well a series of workshops that henning should have colombia and I hosted on behalf of nsf uh on the future of broadband research for the next five years.

A

uh I can talk about it, but the report hasn't been published yet so I'm going to be short, so probably not 20 minutes, but I think it's interesting for this group, because it doesn't actually addresses a lot of the topics we've been discussing and then we're going to talk a little bit about the itf 110 preparation and the fact that we're on the last session of the last day, um this is an official irtf.

A

Interim meeting, so we have to follow the the rules of irtf in terms of ipr and how to conduct ourselves in meetings. So I'm putting that, and so without more talking from myself uh leon. Are you online? I don't see who's online.

C

A

Okay, so uh welcome uh leon, welcome to our group and I will stop sharing so you can start sharing.

C

C

C

C

Okay, can everybody see my screen.

C

Yes, yes, perfectly good! Well, uh first of all, uh thank you for having us here today to speak here. We are the lucky ones to be able to speak here so good day, everyone so uh to do it itu way. We said uh good morning good afternoon, good evening, everyone, depending on where you are so uh here.

C

Let me quickly introduce myself so my name is leon, I'm from rock dam mobile. So, together with me today, we have paul harvey he's our lead research scientist, so we belong to rock dan mobile's innovation studio. So this is just a fancy name for our research and innovation departments for racket mobile.

C

So so thank you again for having us here today. So, let's quickly go to what we want to talk about today, so uh I just want to start by sharing, uh basically my perspective on the authors autonomous network, so so yeah, just just as aeroplanes uh don't take us into space. uh We don't believe that existing technologies alone are enough to achieve the higher layers of autonomy, so so in order, what you see here is basically, where we're trying to give everyone a bit of perspective.

C

So so I mean we will not be able to reach the moon by trying to fly our plane higher.

C

So what we, what we are trying to do here is basically uh basically to see how how we can achieve how we can achieve this uh autonomy so as we as the role of human is reduced in the higher layer. The challenge, the challenges that we face become basically exponentially other. So this is the reason that multiverse motivates us to approach, uh motivates our approach of evolutionary driven autonomy, but we know that we know that we cannot achieve autonomy alone.

C

So that's why we have been forming technical collaborations to work together and research on the necessary ideas and topics and technologies to achieve this, so, which is why we started a new initiative in itu, so the focus group on autonomous network.

C

So I apologize for this boring slide. uh It's a very standard slice with too many words uh I'll explain briefly to everyone the overview, so just a quick background on itu-t so itu is, uh is the international telecommunication union, where the t at the end stands for standardizations.

C

So it's part of the un family, focusing on standards for telecommunication and information, combinatory, communication technologies. So within the family of itot there is a unique structure of group, which is called the focus group, which is the the group that we are talking about here today.

C

So, although uh explain, only i2t is supposed to focus on standard standardizations but to accelerate uh the study and works on specific areas.

C

So pre-standardization works can be done in focus group, so the concept of this focus group uh is meant to be open to all, so it's free for free for everyone to join. So every anyone can join from any organizations whether you're, you're, you're, you're, part of an industry, you're part of operator you're already a part of uh academia university or you can.

C

You can even be an individual expert that is interested to this topic, so anyone can basically join this uh this group, so so a few of us uh point me basically so enthusiastic about pushing this topic forward and creating an ecosystem to work on this topic together. So we pitch this idea to itu somewhere ahead of last year with support from about 30 entities, a mix of industry, academias, some universities, government bodies and we focus group uh on autonomous network.

C

So this focus group fgn will serve as an open collaboration platform for pre-standard study on autonomous network. So this open platform will enable collaborations between experts. Like you all.

B

C

Because different spectrum, you can, you can be from industry, academia or you can be from sdos anywhere, so allow us allowing us to synergize the the right talent, knowledge and experience to see how we can achieve autonomous network together. So what fgn? What what our group will explore and study is. Basically, uh we will study the approaches on some key concepts uh such as uh exploratory evolutions, real-time, responsive experimentations and dynamic adaptations, so I will cover them.

C

So in achieving autonomous network, we are looking to progressively give more control and responsibility to the software so which includes not just operations of the known, but also adaptations to the unknown so to structure our approach to this challenge. uh Our focus group will try and achieve this goal through. Basically, these three key concepts here you see so first one exploratory evolution and then real-time responsive experimentation, nicely dynamic adaptation.

C

So the first one is about typing. In simple words, the first one is about making or creating logic. So the second one is about how we validate it and then the third one is the necessary, tooling and technologies to apply the logic. So I mean logic might not be the right term. uh You can replace logic with solutions like basically creating solutions, validating solution, applying solutions, but but depending on the solution situations, uh I'm happy to hear uh your opinions or suggestions uh here so quickly go into these key concepts.

C

So one of the main questions of autonomous operations is how to adapt to the unknown or unexpected, so for the purpose of basically addressing or improving a situation.

C

So at the moment, such challenge requires a human being to be involved so and to craft a new or better approach using their creativity each time when something comes out. So our challenge here is basically is how to achieve this with minimal human involvement, so uh evolution evolutionary computing represents a codifiable way to achieve uh some form of creativity, so, in other words, evolutionary computing is a way to write down the process of creativity.

C

More concretely is it: is it a way to explore it's a way to explore a large number of options in an efficient way, so evolution basically, is the driver we use to create a new classroom controllers, so how it works is for another discussions, but we are not the only people who are pursuing this line of thinking as of now so google google has recently applied the same approach to enable the rediscovery of several machine learning technologies by a software entity.

C

So what you can see on the right side is basically an example from auto ml0 by google, so the end result of which was a better approach than is currently known to human for a particular problem, so nasa and nasa. Basically, you see the space space shuttle that we are trying to build. Nasa um nasa has used the same evolutionary search, similar average search techniques to design a radio antenna that would basically cheaper, lighter and more effective.

C

Than human equivalence can produce so, but this sounds great, I know uh yeah. It sounds like the the secret sauce and it's like holy grail, but but but there are still a lot of challenge challenge to overcome, so so the questions we are trying to ask ourselves like how how long does it take? uh We don't have a million years to figure it out for for this kind of emergence uh behavior. So what is the right? uh What is the right form of evolutions to use? uh Can evolution be applied to all use cases?

C

Probably not, maybe only, but only few or maybe a handful, but maybe a lot generate cases as well, but but I think we can only figure it out. We can only answer all these questions uh when we try it out yeah for real, so in zipper words, uh evaluation. uh This is about how making or creating logic to improve operations of unknown situations or to adapt to the unknown.

C

So so now that we have the logic, so let's create it. So how can we validate it? So this is where we reach the real-time responsive, experimentations.

C

So so yeah, you can see the interesting visual how humans test their ideas, so so should machines.

C

So if evolution creates a new cluster controllers, for example, so experimentation is, is the way that we ensure that they are fit for for the purpose, either through trial and errors or generally validate them against the tasks they are supposed to perform.

C

So this is also the way that that we can provide relativity and trust in the mechanisms that is generating the controllers itself. So so the experimentation we are talking about is actually a spectrum uh encompasses fractional stuff, so so from sanity checking. So the processor falls process of statically checking controllers, whether they are reasonable, for example, the critical path of this controller. Is it too long? Is it too short? The optimization controller can only turn on a light switch on and off, which makes no sense. If that's the case.

C

So then we talk about the next one. We talk about simulations, uh how we can uh using simulations tracers or test bay to test the correctness or effectiveness of close loop controllers, or that the new controller is better than the current one or not. So, on the other end of spectrum, you have the canary testing you see progressively larger integrations into the real production network. We want to see how we can do that as well and how we want to push this to the real production level as well.

C

So so uh the concept of such testing has already been shown. For example, I'm not sure. Maybe some of you are familiar with the other works from itu focus group on machine learning, so machine learning for 5g, so they they have this framework for sandboxings which uh touch upon this concept as well. So so it's nothing new, it's a it's, nothing, really new or but we're trying to see how we can apply this in in the sense of autonomy.

C

So so so there are still challenges that we are trying to to ask ourselves basically how how to automatically construct experimentations for the merit of controllers.

C

How how and to what extent, can digital twins be used to create experimentation environments, how to effectively effectively create a testing environment that allows us to trust the ai created controllers and what kind of safety measures must be in place to do? Production chemist, testings and then also, I think this we have to ask ourselves at what point is something considered validated so so the trust level of the the things that is being created so yeah. So that brings us to the next one.

C

So so now we have the newly new or create newly created or improved logic. We've validated it. The next questions we ask ourselves now is: how can we apply it? How can we realize it so creating and validating new logics or solutions, but never being able to do anything with it is well, basically, I guess useless, so so, um virtualizations and clouds.

C

You heard this a lot has given a lot of flexibility in architecture and deployments. Well, especially in the telco industry. They are doing a bit of catch up here, so so, augmented by these disaggregations of telco functions, open architectures, the open runs that you see. You heard a lot, so the the openness opportunities now presents uh to how to dynam dynamically adapt architectures to future environments, uh future technologies or future uk use cases.

C

So now, with this opportunity with other things, what other things we are trying to answer is what other things needs to be in place to enable machines to do this instead of humans, so to realize this dynamic adaptation, so so what kind of technologies, what kind of interfaces? What kind of specifications, data models? Ontology taxonomies semantics yeah. So this this? These are the questions we are asking ourselves so so so imagine uh now we have a close look controllers.

C

uh Everyone is trying to build at least at least everyone in the telco is trying to build a ai ml and enabled closed loop controllers. So now, let's take a step back and see how these closed-loop controllers can be decomposed into a smaller function of building blocks. That's modularizations! You can see in the slides so and that's nothing new as well. I think it's been discussed for a long time. So now now take one step uh backward again, one more step backward.

C

So we are asking: how can we basically describe this functional building blocks so that not just for the sake of describing so so we have you see we want to. We ask ourselves how we can describe these blocks so that it can be later consumed for evolutionary compositions, so in the first we'll talk about the creating the logic. So so, how can all these building blocks uh be? Decomposed like like lego blocks and then the the the first, the first actions, the the first concept of the uh the creating logic?

C

How can they pick up all these different different blocks and see how you can compose it and create basically new unknown controllers, just like what uh nasa update of creating a basically out of the world antenna that, basically uh it's just if you, if you google, the picture the antenna shape, is just out of this world that I maybe maybe, at least for me- I would not be able to think of that kind of design. So so this is.

C

This is pretty interesting, so so uh I saw on the agenda as well the presentations by uh aggar from erickson's on semantic descriptors for intelligent services.

C

I think this is also very relevant to this, so how we can describe a lot of different different intelligence services as well uh yeah, so so these are the questions we are trying to ask ourselves and trying to address in the focus groups so also uh existing specification language, uh the data models or taxonomies are these sufficients to realize uh to to realize autonomous networks to realize that the evolutions needed to achieve alternate network so so, based on the the all the mirrors of use cases, so 1001 use cases out there existing interfaces, the protocols, the programming languages?

C

Is it sufficient, rich and expressive enough to achieve autonomous networks? So so yeah it's a lot of questions, uh so we are trying to ask. We don't have all the answers and we don't know what other. Basically, there are new things that is required, but we don't know at this point that we need them at all, uh so so how and yeah these are the all this. Basically how this can be applied to all the use cases at this point. So so this this. These are all the things that we are trying to ask ourselves.

C

So so, with this with this all these questions in mind and this interesting topic in mind.

C

Well, we happily went on to have our first uh meeting uh last week, so uh we received around 35 contributions for the for the focus group first meeting, so so basically contribution means uh input, documents or papers or anything to to discuss various topics that people from different backgrounds they want to discuss about. So the meeting itself we had about uh toronto registrations. We have uh around 100 participants each day we spent lasted for three days.

C

So uh what I want to say uh highlight here is basically the people, the crowd is it's a balanced crowd of people, so so it's not heavily tutored to any specific uh uh spectrum of people. It's like it's not heavily due to the operators. It's not heavily the industry side, no, it's not happily to that just on the research site. So so you can see we have. We have a well-balanced people from operators, industry, government bodies, uh ministry of communications and so off and standardization standardizations uh bodies. So we also have a lot of universities.

C

We have a lot of professors uh coming and talk about. This thought this topic and there's also a lot of research organizations, so so we invited uh irtf's nmrg as well, which gave us a very, very good uh retrospective view of the autonomous network and basically the challenges that they see that's the hit. I think this is very, uh very good for the group, so so so out of this first meeting, we we formed seriously try to help.

C

We have a work group structure to discuss uh all these questions that we have in mind in a more structured way. uh I will not deep dive into this today. uh It's a bit boring part of this is so but yeah. We can discuss this.

C

Furthermore, if you're interested so yeah a lot of these questions that that I ask is, are basically research questions so and- and we are here today in this meeting, to ask everyone here in this meeting uh everyone here- if if any of these uh strikes are caught, if you think uh you are working on topic related to us, we are happy to sit down and discuss basically discuss more and hopefully hopefully maybe you can help us to answer some of these questions.

C

So yep, that's yeah. uh So for the standard closing for more information, uh please visit our website. uh You can join our mailing list uh ins from from our website. uh You we have a mailing list, so you can join the mailing list uh uh to to to to join our discussions.

C

We have weekly meetings as well the the agenda, the the timing for the weekly meeting is there if you have any questions, feel free to contact us talk to us anytime, oh yeah, so this uh I just say we we hope uh not not not from just raqqa mobile, but we from the focus group itself. We hope to collaborate and learn more from coin rj in the future.

C

So that's all from my side. Thank you very much.

A

um Thank you um looks like we have a question from dirk. Yes, actually I was going to say uh a lot of time for questions.

B

Okay, um yeah thanks a lot leon for um your presentation.

B

um I um have one question so as far as I remember um so in autonomous networking or self-organized, networking has been around for quite some time, so I remember, for example, in the phase where the research for lte was done so self-organized networking for interference management, these kind of things, and so I wanted a bit. um What is um now so I knew from your perspective.

B

So at that time it was almost very um always seen as a kind of management uh approach, and um so I'm I'm kind of wondering what is the new research in your point of view, and so do you see any connection to um the computing in the network um topic that um this loop is uh researching.

C

Yeah, thank you.

B

Yes, it's nothing new. We should.

C

B

C

B

Nothing new, I'm just wondering. So what is so, I guess maybe machine, learning or sdn could be new elements, but just wondering.

C

No, no, you are, you are, as you are exactly correct. It's nothing new, don't tell the telco peoples but yeah, it's nothing new, but but but the thing is uh like I say so: the telco we from at least from from an operator point of view. So so we did a lot of catch up. So uh what happened here is basically uh I'm not trying to boost here, but basically uh roclet mobile.

C

We we managed to deploy uh it was I don't know how about we managed to deploy a sort of end-to-end cloud native uh telco network, so this basically unblocks a lot of new opportunities, so alternate network in computing itself is not nothing new but but in telco itself. This um this this. This creates new opportunities to to look at this area, so so so, basically open up new new doors.

C

How we can achieve this in telco and basically yeah uh it's nothing new, but it's something something that something is something new for for the for the telco industry and- and this is something that we can actually. uh I think that the telco telco industry can actually benefit a lot from this. uh Considering the the massive amount of uh the massive amount of devices and things we need to to manage now with 5g, taking the place iots and everything so so, like I said it's nothing new, but to from a telcos perspective.

C

This is something that can actually bring us brings a lot of benefits to the industry.

B

D

Other questions, uh I have a question. This is eve schuler from intel, and um I guess my question leon. First. Thank you as well for a very interesting talk.

D

I know that you said that you don't think that there's anything new, but um when I you know wearing the the glasses that are, you know through the lens of coin rg. I wonder if you know speaking to your point that maybe this is an introduction to the telcos.

D

It raises the question of given the networks and the infrastructure that exists already. um If the deployment of the compute and the I mean. So we you didn't talk that much about the compute.

D

You talked kind of indirectly about the composition, which I do think has some relevance to um the activities going on in the 1dm world, where there's a working group in the itf called astl a s karsten, please correct me, but the a semantic definition format asdf, um which does look at a language if you will for how things interact with each other, and so it would be interesting to understand how those sorts of function blocks come together.

D

But, additionally, some of those properties of those objects are the kinds of operations they support in their interactions, and and for that I do think that there may be um not just you know what operation should support but where to run those operations, given that there are many that the location of what it is you're trying to do may not be on a single host that it's potentially across many collections of resources.

D

So I know I'm rambling here a bit, but I um I'm hoping that the conversation going forward um in the new itu group um will comprehend that kind of discussion. And so you have have you seen hints of that or have you seen organization, other organizations that are touching on those kinds of questions about where to run the compute? That's going to help with the evolution and the composition um that it isn't just a centralized activity.

C

Thank you eve. uh Well, uh if I understand correctly, yes, uh we have collaborations uh basically on edge computing, so we have. uh We have people from universities trying to working on the distributed computing as well on this so again,.

C

Again, uh if you think this is, I.

D

Know it's very new. We will join in.

C

Yeah well again, uh I don't. I don't, have all the answers. So if you think this is something worth pursuing uh in in terms of work, I mean this. This is the something that uh we're pursuing together.

C

We are happy to sit down and see how we can work out on this area. So this sounds interesting so but yeah.

D

That absolutely does thank you for coming in and sharing yeah.

A

I was I was going to say: I was going to say that I think that aspect would be a common interest for um both uh coin rg and the itu group, and I think, we've agreed to to talk about it so and we should continue um so we're we're now going into.

A

um Are there any other questions or should we move to the next presentation?

A

uh Other questions? You can go to the list um leon and your colleague wait. Don't remember the name. uh Please join the the failing list, the mailing list, because if there's other questions it could go there and there could be uh more discussion on the list also about how these two initiatives.

A

Overlap and could eventually um uh yeah collaborate or actually move together and towards a few common uh common goals. uh Research goals.

E

Sorry, uh paul I'm the other guy with leon. Thank you very much for having us and just to echo what leon's comments were. I completely think that there's overlap between these two areas.

E

Well, we do have some efforts on edge computing, the compute and the network, especially how to take advantage of the existing infrastructure and all the necessary mechanisms you know about the mechanisms to achieve that is a strong overlap, not just as was pointed out by eve for where to run these evolutionary computings or any other approach we use, but also for the use cases and services in the network themselves, yeah so yeah, a hundred percent.

A

Well, it- and I think it goes back to uh some of you- may know the uh the eu's also looking at creating uh operating systems for iot distributed networks- and uh you know, we've had this idea that a network now is much more like a a computer board than a uh a series of telephone interconnections.

A

So um yeah, I think, there's going to be uh potential cool work to be done. uh Moving on. Actually it's interesting because the next presentation is from alex and uh leon uh started. Some of the things saying: there's nothing new. um Well, I think there are things new, but alex will tell us what is not new and what could be new so alex. Please start your presentation.

F

Hello, can you see my presentation.

D

F

Very good, okay, my name is alex callis, I'm from university college, london, obviously in london, uk and thanks a lot for this opportunity to make some presentation about this important topic of programmability.

F

It just happened that I have published more or the same number of papers, research papers in programmability and autonomousities. If I will try to make more links between the two area. However, just to study autonomicity is ten years younger than programmability in networks. It's 1971 the first good paper all right.

F

Both these areas, I have to say they didn't reach yet the maturity which is necessary to go into commercial networks yet, but hopefully the next period will allow that to do it, for my presentation is mainly to draw some lessons from the past research and engineering activities and to avoid pitfalls for the future period, as applies mainly to programmability, but also to autonomicity.

F

um I have a substantial number of content which I'm going to go very quickly through and not to deal with in too many topics. However, I'll concentrate on some suggestions about what could be new and important for the future.

F

I will go through uh some sort of content list, starting with very important the key con concepts, the terms used for the last 40 years in this field and I'll conclude with some suggestions for the future. Now the key terms in this area are related to programmability, which is in simple terms. Is uh it refers to executable code, which is activated or injected in some sort of computing execution environment in network elements of different kind?

F

It could be any network elements in the wireless or wire line, access, core edge and even space, which means that it's quite a substantial type of facility which allow new network functions to be activated on demand and managed and changed on demand.

F

So it's a very powerful technique.

F

uh However, it has to be anchored in a lot of security environment in order not to allow this is such a code to create havoc or problems to the networking in general or in particular, to some parts of it in any form of shape. So the two parts are absolutely hand behind. No one will use programmability if their security are not extremely highly defined architecturally, as well as implementation. Wise, very importantly, to look at the concept of in-network computing and programmability is the fact that it allows definitely the network elements to dynamically be changed.

F

However, this also raise opportunity to link it with a network services either one way or the other way in such a way that basically making the network services being network aware and the other way around. Something which is important in these days to minimize enormous cost which exists in terms of dealing with uh with these subjects for the operators and others.

F

And these are basically the terms used, maybe different uh terminology- I I looked at the coin terms and also it's their line of this part and we can move forward now. There are a number of lessons already, which I want to draw your attention.

F

In the past network, programmability were also hand-in-hand with developing relatively new platforms.

F

Many software platforms. This is quite expensive to do tricky, to organize and even more complicated, to create a big user group for it.

F

As such, it's a time to look at the programmability to in terms of piggybacking, on the massive trend now which is softwarization, which is happening irrespective of programmability or economicity.

F

A lot of new platforms are dealing with components and system issues at the software level and obviously programmability or and analysis about software. So it's an opportunity to piggyback. In this sense, that's the first lesson take advantage. There is no need to invent massive platforms which an operator will accept, and I'm pretty sure that you understand that this is quite a big undertaking and there is no need to do it, provided that programmability and autonomicity in particular. These two elements could be added and tested an existing, let's say successful platform.

F

The second very important lesson from the past is that sometimes it's good to not to look at the problem from the past, because by the time you solve them, they're already the solutions are obsolete. The situation is moving.

F

Currently, the industry is moving towards other next generation networks, in particular, two types which are interrelated, which is network 2030, as it is specifying details as architecture level by itu-t or a subset of it is 6g, a wire line, part which is also fully architected so far in proposition terms with white papers from different different sources of magnitude.

F

But clearly these are new networks where the requirements are different and some of these are listed here, which involves stringent delivery for kpis guarantees of different kind, autonomicity, programmability and many others, and so this is an opportunity to to look at these problems related to the specification of education networks, in particular.

F

The two which I mentioned, which are overlapping anyway 6g and network 2030, and concentrate on this as a way to move forward the last, but not the least, actually probably the most important lesson from the past is the fact that.

F

Programmability in particular requires a change in the network devices and as such, if you propose something to massively change the current devices, the cost of doing that is prohibited and therefore, as a result, any solution will not be usable. So it's a time to look always for a delivery mechanism for such programmabilities. So it's important to do it.

F

One of the systemic way to is now everywhere is about slicing, which is now fully researched and a lot of system organized and in such a way that programmability, based on a slice, has a better chance to be deployed. Bearing in mind that another slice there could be other techniques applied and different, and so on, and in this way you don't assume that it will be deployed at large, which is impossible.

F

Another that's the three lessons which are applicable to both this area: autonomicity and and programmability.

F

Now, in the past, in programmability terms, there were tons of research results. The bottom corner from the left includes, let's say, activities where millions of papers were produced, for example a routine packet forwarding, but also on the other side, node operating systems, programming languages and so on.

F

The blue centers are more or less relatively new, 10 years or so 5 years or so new models, which tend to help the process on one side on a computation side, you have network level operating system facilities and continuous from center cloud to edge cloud compute.

F

On the other side on communication, you have network functions, control, slicing control, kpi controls and automatic control, but they are converging and they're moving. There are a lot of results on this blue part, a lot which I'm going to go through quickly.

F

But one thing which I would like to stress is that there is no single model or a single solution which hits all, and that applies to both this area. We have to be careful to look at other things which will be interworkings as well. Now.

F

What are the key results, mainly from standard groups, one of the early one which uh quite early in is called project p 1520, I travelly 20 and it's more than 20 years old, but it's the first one which defines a way in which precursor of the existing today devops autonomic devops, which is applied to network or service chaining. But in principle it created a suggestion. A number of interfaces by which network services could be composed from pre-compiled building blocks, which are composed in you, know particular logic and then deployed.

F

So the idea is more or less the same like uh in today's devops, but the original one is iso, and this is a substantial way to move forward on this subject, even a level of programmability the second one. Very importantly, it's an iltf forces group, which I notice that van gaalios is on the line.

F

Just was one of the people waiting on this, and that group is relatively old as well nearly 20 years old, but it reached a level of success in terms of proposing a separation between forwarding and control elements in the network elements, and this is now, in my view, applied to most, if not all the current network devices, although not using the protocols and methods proposed by forces, but it helped the lower level environment. If you want like this to move forward in terms of what it need to be looked at another.

F

Probably the most prominent area in this field is about related to active networks or program networks and the scenario which started in united states, but it was quickly covering the whole world and more or less suggesting that there is a need for decoupling, network service and applications through network apis from the node apis of the node devices and to use some execution environment a place where you you execute computation practically into it, and that could be very near to the node or very far from the node.

F

It doesn't matter but to allow this to allow a level of programmability. So I did quite a lot of work in this field. There were more than 50 big platforms available for experimentations in this area.

F

Last, but not the least important uh standard result is about sem. Probably the most important part of sdn is about the fact that it allows a level of programmability of the flows.

F

In the network, currently, in my view as a whole, is not necessarily very successful, it's always dying and will be replaced over all those other scenes, but at least this should be mentioned now. If you look to the whole scenario of all the current research results, hopefully I didn't miss too many.

F

I put them into one slide.

F

They contain more than 50 separate elements, presenting research with multiple risks. Research paper published experimentation, platforms, developed and tested.

F

If you want and many more and they are related to different parts of the way in which programmatic could be looked at, and I think that all of them are important, and that also also proves that one solution doesn't solve all and they are related to adds on to the let's say the operating systems either to the ones which are open source or the one which are proprietary data plane devices in particular many such things and execution environments, different type of execution values.

F

There is no single way where computation could be attached either and placed near the node or very far from the node at the bottom. You have a little schematic about potential.

F

Let's say blueprint for programmer devices and which could be applied to access core and edge, in particular, a variety of possibilities. However, some of this- let's say research survey in detail, were part of a book which was dedicated to programmability on network, which I published some times ago. I updated this slide based on other things, but uh just a little bit of warning. That book has around 600 pages.

F

It's a research book I have a 60 or more pages are dedicated to security, are showing that this is absolutely essential to offer a level of confidence to anybody who wants to program networks and to allow the operators to do it.

F

Sometimes this is over overlooking autonomousity is the same applies to the same subject now. What are my proposals.

A

Alex um do you have still a lot of, um because I think this this could be a a full day tutorial and we could think about it. Do you still have a lot of material? um I.

F

Have two slides left? Okay, so what would be my suggestion about taking this into account, and maybe that could be taken further based on the four?

F

Let's say, lessons learned from the past number one: it's still time to look at forwarding playing programmability and which is more or less an attempt to re-engineer the ip layers, in particular the user-defined networking, which allows the packet header to be changed to allow programmability from the user point of view to happen, and this is clearly needed for deterministic coding, as well as attempting to replace best effort with some level of high precision networking services, as required by the next networks.

F

Another group of management related uh challenges for this programity is listed here. I'm not going to go through it, but a few things which are important, triggering programmability in a network is as important as a network is so the two should be hand in hand both will create problems or solutions, and without them together they will have massive massive issues now.

F

My last suggestion is to look also how to apply to provide these a multi-domain issues, given the fact that the current networks are multi-domains, mining operators and within's an operator have multiple domains, so this is absolutely essential and this scheme, this diagram, shows that programmatic and autonomousity could be put in helping these multi-domain issues to be resolved.

F

Now what has my conclusion remarks, which I'm not going to put in place too much, but one of it is important. Is it's a time to move, involve your network, computation programming techniques.

F

To solve problems which are now identifying new requirements in the next generation of networks as a way to be accepted. Obviously there are some benefits which are described here and also it will help the network services to explore. So that concludes my short presentation, which I hope that also links with the previous question from the previous presenter.

F

Well, thank you.

A

Thank you. Thank you very much. um I I it's time for maybe a few questions. If not nobody else has. When I do, um I thought well, I saw you were mentioning that net 2030, which has been uh mentioned in in many irtf uh forum before, and I think everybody uh took notice of it.

A

uh My question, however, is that I think that and- and I hope the rakuten people are still online- I will paul is there is um what one thing we tried to do in coin is not have that much of a operator focused um and look at it more as a functionality of networking in general and also to encourage what I call the new operators, uh the people like rakuten, who come from the uh the data centers and the uh the cloud infrastructure more than the traditional uh operator in infrastructure.

A

um It was also one of my criticism from that 2030 by the way- and I understood why, in the context of net 2030, but I think us we are not uh linked to uh any operator or any company, and I would like to have your opinion alex uh how you could you know.

A

I will put your slide as as interesting and very good tutorial background, but how do would you see this evolution again when you stop thinking of the traditional operator network, but much more, a number of processing nodes that are connected and are used to really really deal with this new data-driven world we're living in.

F

Hi. Thank you very much for the second answer. It's already answered in my presentation. If you pay attention but I'll repeat it, a lot of activities done in the last 40 years and in particular last period with a view of how to improve programmability for clouds computing and so on. The techniques are excellent, but.

F

At the end of the day, the proof is in a pudding, namely the techniques that we used for real or realistic network developments, and it doesn't matter who is going to use them, but they need to be good, and a lot of waste, of good ideas on platforms were with the fact that they didn't think in terms of the fact that they need to be used for us by use by the network.

F

Designers and developers of new network functions, also cloud functions and to be put in place by operator to be managed properly. Without this full part, some technology developments are useful, but has only a small longevity for the warring is each if you want to put it simple. Each of the coin.

F

Drafts or activities should also cover explicit, is a one which should be deployed easily and use, presumably in a network or cloud just a new technology or new research and technology is just the first step, which may not be sufficient to convince other people to move forward. Therefore, that's basically the answer for network 2030 was not used only from the operating point of view. It's also a list of new characteristics which need to be achieved, but of all the same applies to 6g networking, which I'm I'm also starting to understand better.

F

As I participated to so network architecture, description of the 6g, probably the most the best in town.

F

Flagship and they are basically needed, new characteristics which cannot are not available today, so these are the problems to be attacked. First, we should have a better chance to be successful to be taken by the full life cycle of the development of network functions.

A

uh Thank you report this for this um answer. uh Yes, it's very funny uh edgar who's on that call, and I were talking about it a few days ago uh paul you have a question.

E

Thank you very much. Thank you alex for an interesting presentation. My background is actually programmability for heterogeneous distributed systems, so it was really nice to see a lot of familiar names and things popping up, um and I think I agree and then to blend that with the chairs comment to bring this to a conclusive point. I think it's an interesting time. uh We've all heard about the virtualization of all the different networking functionality to turn telcos into typical clouds, and at least the way waxton's looking at it is quite like the sharepoint is.

E

It is a collection of distributed system or distributed elements bringing together to operate the network, with the caveat that you still need to provide all the connectivity that's happening, but as you move more towards software-based systems, all these interesting questions that are being asked is that what is the right way to represent it? What is good research, and how can we use this to effectively operate? Of course, I speak from the operator perspective not to plug it too much, but I think that's what leon was getting about is like in the autonomous network focus group.

E

That's the place where we're looking to try to ask these questions and because the truth is no one knows right. We talk about autonomous networks. No one knows really what it means and it's an opportunity to bring all the different perspectives to the table. Put it on in front of everyone and say well what makes sense. Well, it doesn't make sense and what are the wonderful ideas from research we can take and apply and what are the terrible ones that we throw away so yeah.

E

So I think this idea of how do we program these things in the right way for the domain in which they're supposed to operate is a very important part of that conversation.

A

Okay, thank you. um We will uh move other questions to the the list. If there are some uh the next presenter um is. I I remember my um agenda so now um I'm loading it okay. Next one is alex. I was wondering I think we kept. We kept the best, the best the best for the last now I I thank you rajasthan for your your patience. uh Next one is edgar from uh erickson and again, like um um leon was, was mentioning.

A

uh This could be some link to what he's doing uh for people who are asking. uh What is the goal of today's presentation?

A

uh Presentations, they're they're a little bit, I would say a dusting of a ton of different topics that were addressed uh on the list and in previous um uh presentations and uh in previous meetings, and I think it was the start of this new model modality that we want to have, which is more focused on papers, presentation and less on drafts, and uh we will continue that at the at the ietf as it's coming, and I think you have to to bear with us that we're starting this new model.

A

So, uh of course, we're going to have to you know, tune it to the um audience so edgar um again uh for people. It is not I'm the co-author of that paper. It's not exactly about computing in a network, but it's it presents something that is very important for computing in a network which is finding what you need and finding it in a way that you can actually share it with other people and using a common language to essentially be able to share that information.

A

So edgar, please.

G

Yes, thank you jose. Let me see share my screen.

C

G

You can see now my slides.

A

Yes, thank you.

G

So I guess, as uh mary jose has said, this was a part of her work that I have been doing together with her and also a couple of other authors, timoth schneider from the university of alto, which is now in in the netherlands, back and being the master who is uh from the again university of belgium who had worked a lot also with semantic descriptions of functions.

G

So then the idea was like: um how can we actually do something that could serve uh also for the iot uh or not only try to um do something that that it's only for let's say, machine learning in general in cloud? So basically, this is where the networking aspect comes into play. So since there is iot devices, so the iot device is going to be connected. So what I'm going to talk is a little bit about.

G

First of all, what are these services and why they? They are important for iot, then second, how we came out with this kind of taxonomy of describing these services and then finally, I will go very fast towards one one example: using a very well-known convolution network called mobinet.

G

So, first of all, some introduction is about.

G

We have been talking that in order modify or or let's say, handle intelligence prices, it would be needed not only in devices I mean in infrastructure in general, in the edge, so we would need to to add some sort of abstraction that you can. Actually you could actually operate to so that attract abstraction would have their own layer today that we found it together with the application.

G

So, basically, every application have an api, have some services, and then it might have a machine learning model, for example, and then that all is packed in an application they down down thing of this is that if you want to reuse that model, let's call it a system, a model for a speech recognition and you have other applications in the same device or in the same system. You basically cannot reuse it, because unless that application has an api that explicitly allows you to use that that model, then you cannot actually use it for something else.

G

The other thing is that it's very tied to how the application is built. It means that interoperability is an issue. So if you want to actually change suddenly something from that intelligence, you cannot do it, because it is so much embed in that application that you cannot really modify it.

G

So then an idea is that, can we actually create this abstraction layer that we can call it um intelligence layer if we want- or we have been talking here in coin, before of about the the data, the data layer? uh Our I mean we can't give it any name, but the idea is that this layer would be basically very similar to an operative system. You, when you um are writing an application.

G

You don't think in the application how to send a packet over tcp you basic line of code that allows you to send that rig to the operating system and the operating system take care of delivering those packets or that object. uh It will do the segmentation. It will do everything for you so why we cannot do something similar here in the devices where we actually abstract this layer of intelligence, where we have all these services, like recognition, prediction, anomaly, detection or even generation of synthetic data that the applications could ask hey.

G

I want some uh recognition of of the camera image that I have here and then you get an answer. This is a an elephant or this is a traffic sign and so on. So then, this is the idea. The intelligent services would be these kind of exposed models, ai models that would be running in this kind of abstraction layer called intelligence layer now.

G

But if we think about this kind of model, when we are saying that we put things in the devices and then the devices become suddenly italian, we have to think about um the context that that the iot has as such, with when he's trying to to address a use case, and also, we have to think about iot as a device, a heterogeneity as such and then also the context where actually you want to solve or how you want to solve this problem.

G

So if we start with the problem solving, so let's say that you want to do a recognition of something so, for example, an apple, so that would means that you will have basically different use cases where you could use this kind of recognition.

G

Each of the use cases will have their own requirements, so if you have a for agriculture most probably you want to know if this apple is sick, or is it actually healthy for the suplate supply chain? You want to know if this is really an apple or it's an orange. You didn't want to know anything else, but that you can sort it in the right place and for diagnostics. uh It might be that you want to know if this is um growing.

G

Let's say accordingly to the um to the standards that were supposed to to have this kind of apple.

G

Then another problem that you have is where you are running this software, so it might be that you might be running it in the data center that you don't have any constraints of hardware, but it might be that you have latency problems if you are running it in in the age and then you need to actually bring the results, so you might have to move that model closer to where the act is happening, for example the brakes of your car.

G

You wouldn't like that the decision is going to a data center somewhere in the u.s when you are driving in germany and then deciding if you should stop or not. You will like that that processing happens in your car and also the different type of processing environments might be more complicated or less complicated according to what is the capabilities of those devices. And finally, the solution context is like: how does it look the algorithm or whatever you are using in order to to do this recognition? So can you recognize a apple from the smell?

G

Probably, yes, can you recognize an apple from the picture or from the shape or from just from the color? So then the idea is that the solution space is also quite wide, so you could actually have solutions.

G

That is not only a picture like with what you would expect, um then the the the thing is that why you want to have this is because, when you are creating, let's say um systems like a smart home where you are having an application developer, building applications on top of your system and when we are talking about a home is something that is expected to last 100 years.

G

And then you have a speech, recognition service for that, and then imagine that can you I don't know, change that system depending of the uh what is needed for the persons that are going to live there. It might be that I already have train and speech recognition system why? I cannot use that one instead of the one that the maybe the application developer has a deal with. Can I interoperate that service and say, like you know, instead of having this I don't know marcus.

G

I would like to use billy because billy is, I have been, I have it in all my devices and- and it really understand my english, so it's perfect for me, or it might be, that you are maria and you want to have alex, because it has an excellent spanish version of speech. Recognition.

G

Also, I mean for the for the application developers themselves- it's quite nice because they don't have to care what would be this service uh um service that is being connected with your software, but then this connection, how you actually map what the person is talking to what the system should do. So that is a semantic problem and there is where we are trying to to address it with this solution.

G

um Another thing is, as I was saying, there are software dependencies from the solution so many times they they have been. um They have been written in some programming language or they have an execution environment that the one for the model that that is provisioned to you and then also there could be hardware dependencies due to either acceleration, because you will require this kind of um more specific soft hardware.

G

For for doing the the instructions, um execution faster or then memory or then sometimes also even network connectivity, then another thing to take in account is the policies not the whole story of intelligence comes from what the device supposed to do, but also what are the business relationships behind those and then it might be that you have certain requirements from security from networking and from performance.

G

Other requirements like well. I want my bananas to be 50 degrees, the whole time they have been in the logistic chain, so how these kind of policies can populate all the way down to the devices so that they can understand what part of that policy corresponds to me and that's another topic of research on looking how these semantics should be expressed and then populated all the way down the stack.

G

um Then, if we talk about intelligent service, we can think of intelligent services as having data. That is input. It is processed like any any any system, and then it is actually output. Then this input we can think about multiple contexts that this input can have. So let's say that one content, the data is coming from the device.

G

Another context could be that the data is being feeded by an external system or the data could be fetch from a pop subscribe system and the same thing happened with the output. The output could be that well. um This is a result that I'm getting every, um I don't know every 50 milliseconds and then this the other stream might be results that are consolidated over one day.

G

So then, you could have actually multiple meetings of the same data that you, your system is capable of, producing and the same thing, the meaning from the data that is being in input.

G

Then another thing is that we could define in these streams of data, something that we call tensors in the machine. Learning world is well known. This word tensor, which is basically trying to address a multiple domain that one data object could have so one one example is a picture. So if you have a picture, you will have a rgb uh mattress and those will be atomic objects, so you will hit.

G

You will have the the the blue, the red and the and the um green um values, and then they are together with, let's say the picture. um What is the resolution, and then they have a mattress that that make compose this kind of data and then also in the output. You might have similar things, for example, for machine learning functions. You might have a classification mattress coming out of that of that intelligence service.

G

So the idea we have with these um semantic descriptors is: how can we how we can describe data and describe these services that is being input in one hand and also what is supposed to be doing so? What I'm presenting? You is basically how we have thought that this data input and output could be uh described, semantically in a structure, then this has a total. I mean total one-to-one match to any of the data models that already exist.

G

So basically you could actually map any of the data models to something that could fit into this kind of uh format. So it's not something perpendicular to anything is something that can be adapted and also they might be domain ontologies and vocabularies and even custom things that that can't be adapted in this kind of data structure, and the idea is that whenever you get these services, you can use them to create pipelines.

G

So you can take these services and then put them together and create more complex services, and these more complex services might be even customized to a certain transformation so that you can serve other services or even store the data. For example, then these transformation functions could be also something that resembles a service, but it will be much more simple, so it will be something that basically do a service for the service that is is making or preparing the data so that the service can be used.

G

um This is what we come as a taxonomy here, of course it looks like it's a lot of stuff, but then what I want to just highlight from here is maybe the the left part where mostly all of the things are, are being instantiated as some kind of domain, specific description in ontology, which means that you can describe the service in the same way but using multiple domains. So you could have an ontology that it specified things.

G

For um I don't know uh gardening, and then you could have a vocabulary for gardening and then you could actually explain that service from the garden in perspective. It is a service that applies for that, but then it might be that you have a similar ontology that that could describe things uh for a data center and then there you can actually do this kind of abstractions from a data center perspective, but the same service could be described then use it that different domain. And then here we have multiple um entities.

G

I I will start, maybe with the abstract classes uh very fast. So then this was what I just talked about the domain specific and then they are. These unique classes, so that's something that cannot be repeat so in some of the services you can have. This can can have this kind of many domains and there are certain things which are unique.

G

Some of these things that are unique is who made this service and also it another. Another thing that is unique is: is the implementation?

G

It's missing some some letters here, but then the idea is that this um service has a version, and then this version solves one or many problems, and then these problems are also domain specific, but if the problem is described mainly as a goal, so what is the goal of this service? What it tried to to to solve- and it might be- that is a apple recognition for somebody else is fruit. Recognition for somebody else is just object, recognition, but then what is the actual objects that are being recognized? That's a different thing, um then.

G

uh Here I am mentioning policies, although we didn't do work in policies yet, but we think that the the the server will have in the future. This kind of descriptive policies that up to this service and also maybe to the data that the services is being processing, but then this is for future work. uh The other part is the implementation, and implementation has a lot of things. So here is where we see the data structure.

G

Part where you have the streams, then you have the data, the data tensors and then the data objects, and then all of them has a data label which can be annotated in any domain. Another thing we have is the data sets, because the data set is data that not necessarily is being input or output, but is necessary to understand that data service, for example, training data. What was the data set that was used for training, a machine learning model, or what is the data set?

G

Use it for the mapping of the post processing of that of that and services is the classification uh um algorithm. Then, if there are some dependencies, so it is needed if you require any execution environment or you require a specific hardware and then here we have the pipeline, which basically is a composition of uh these streams and and and and other intelligent services. And then we call something atomic service to those services that can be can be um orchestrated in the same pipeline together.

G

So I will go.

A

Far edgar um are you at the end, because you're way over again.

G

I'm over time already. Okay, then let me just show that we did something similar like that with this model mobile net, which what it does is basically a object recognition and we um I'm not gonna, go through the details, but basically uh we we were able to um delete and then have some, um for example, domain specific uh descriptors.

G

They are machine, learning, neural network methods here that were used for the implementation description, for example, and then also there were certain annotations for the data data labels like, for example, how to describe the the rgb channels uh using um just basic rdf and then some some uh other domains like exif that are used for gpec, for example, and um then we also put on top of that a pre preprocessing, uh a preprocessing function that was basically doing a normalization of the data we added to the model with the pipeline with the pipeline feature, but well to conclude.

G

So the idea of this semantic description is is basically trying to provide an um some kind of structure so and describe this intelligent services in a way that can be interoperably it. The devices need to have the full domains, the everything all the whole information, but the platforms which can have access to any kind of database and or any kind of domain. Specific information could actually do this translation and then send to the devices what they can understand.

G

The same thing: uh it provides possibilities to extend these policies and then also the data set information to understand from where the decisions has been taken, how the things are being operated, and then this is one step just in this automation of intelligent services, so try to get machines to talk to machines, understand each other.

G

So that is what I have sorry for the overtime. Thank you.

A

Interesting, thank you very much. I I think, there's um there's a few questions and I think carson had a comment and I think.

A

So uh we maybe we can start with the question which is again from paul and then we can uh move to uh the comment from carson and maybe the comment from dirk and phil paul. You can ask your question.

E

Very much thank you for the very interesting presentation. um I think there's a lot of alignment with some of the things leon was talking about before my question is basically uh so: okay introducing an abstraction layer to represent all the different data sets that you have is super important have, but one of the key purposes of doing it is for reuse. So I saw in the examples you gave you showed how it could be applied to different areas, but have you tried to reuse those semantic descriptions for like a secondary domain after the initial example?

E

G

um Not yet, but uh I have been thinking of examples of that, uh but we haven't experimented yet with that uh secondary domains, but we have uh doing panama charges. That's what I can say.

E

Cool that would be interesting to know more thanks.

A

Okay, uh carson well carson. You want to talk about your question. I think uh carson and turk are both in the same direction. That girl is that. um How is that?

A

It's not really the network, but how can the network help here.

G

Yeah- and I think it was a bit into what I was saying um in the end- one of the things is like the network can be this kind of translator between domains, so you might have a refrigerator which have a domain, a specific description of what the server I mean, what what is the refrigerator and what services this refrigerator requires.

G

The the refrigerator will have an application. This might be, I don't know, uh planning the food of the week for the family and then in the in this planning of the food, there will be multiple steps which are basically a pipeline of uh different services and atomic services. So then, how you are, first of all acquire those models, uh how you actually get the data for um for a process, not only the data that the refrigerator has, but also most probably from other devices that are over the house.

G

So then, how you actually get the data to to be processed in your refrigerator, so it has to be forwarded from somewhere, and so all those things are services that the network could provide the discovery where I can find them find this, and then you give your requirements. This is the type of data I can.

G

I can um give or I need, and then, where is the service that I can use that that accomplished that for me from the point of view of the player, so I think that those are things and, of course the pipelining thing is inherent something that could go over devices. Not not necessarily the service might be in one device, but I can think of certain pipelines that could go to the, for example cloud and come back and then continue in the device. For example,.

A

Okay, uh we're out of time, and we still have one presentation so thank you very much.

H

And alex raised his hand.

A

H

I'm sorry alex.

A

uh Yes alex sorry, I missed you.

F

Yes, very good presentation, interesting one comment: the main diagram defining the architecture where you introduce a new network services is sound and good. However, there are missing elements here in order to make it usable number. One is adaptation to the lower level environments, because you'll be many types of devices numbers.

F

Two very importantly, some sort of management function to allow new network or many services to be added or new functionality, be added or changed, without which the whole thing is static, uh and it has to be a show how it expands and also, very importantly, to distribute to to define where this network service resides. Is it in a network of mini clouds or it is in uh in one cloud or both?

F

These are the issues which need to be looked at, but the question is very simple: there are eight to my knowledge, standardized, fully developed data models which have also mapping between themselves. One of it is from itf group so for the issue is not to develop another data model, because this is going to be more difficult than you expected in order for acceptance looks to me that you already said that this is not the same is similar to others. What is needed?

F

What- and this is my question- are the information systems which enable data to be collected, distributed and consumed by the network services? That's where it's missing, which is not part of a data model. It's about the use of data model for a purpose, and that's where please provide some suggestion. What to do, because currently it doesn't show it another data model does not do too much. To be honest, although it's similar to the other.

G

I mean you, you put several things here in context and the one thing that is important to have in mind is not a data model. It's a taxonomy to annotate data models, so you will need data models to connect them to this. What it tried to do is to bring out a description of a service and as a description of the server you you mentioned, you need to have a, I mean the environment where it's executed. We try to bring it to the highest level possible and then that might be that description itself it.

G

It might be domain specific. So there might be a description of a refrigerator in the refrigerator domain, but then might be somebody to actually describe that hardware from the perspective of the actual microcontrollers and then you could have very equivalent representations.

G

So what we are trying to do is like actually trying to make possible that you can use whatever is preferred for you and use the network as the facilitator to find what is the best or what are the best services that match so that you can do life cycle management of those services. So exactly what you say: you want to update a service so how you update it, uh I'm finding right match.

A

I will stop this this, actually, maybe we can take again. We can take that discussion to the list because, frankly, uh we have rather stint who's, been waiting and would really like.

B

A

To to present so radish, then um please, uh okay, so rajasthan is presenting a paper that he uh already has uh published in a conference, and uh it is very much after about the implementation of computing on switches and in that case multi-tenant. So now we're going to we're moving from data abstractions into really hardware.

A

uh So please, rather than.

D

Radistan, um am I correct to say that you won the best paper award for this paper.

I

um I'm not aware of that, but.

D

Okay, maybe I'm confusing it with something else. Okay, it's pretty cool, sorry to interrupt.

I

It it was published about a month ago, at the beginning of december.

A

And again um I I really I don't know noah is not online, but uh thank you very much to noah for having uh introduced this work to us. So please, uh rather than please please present.

I

Thank you um I'll try to share my screen and see if that works.

I

um Can you see my screen.

A

That's perfectly.

I

Okay, thank you um so uh hello, everyone. um First, I would like to thank the organizers for inviting me here today and for giving me this opportunity to present our work on multitank programmable switches.

I

um My name is razden stubyanov and I'm a recent and film in advanced computer science graduate from university of cambridge and today we'll be discussing how multi-tenancy can be supported on programmable network devices.

I

In recent years, many applications have been developed for in-network computing. Some examples include inbound network gallery used to collect and report network state consensus protocols, significantly increasing throughput and reducing latency stateful load, balancers capable of replacing hundreds of software instances with single switch key value stores, providing high throughput and low latency with rapidly changing workloads.

I

And machine learning models used for in-network classification.

I

This type of network applications are particularly useful because offloading software functionality to the network allows us to make available cpu cycles to user applications.

I

However, today data centers are shared by many organizations that don't necessarily have the same requirements in terms of availability and security, and what, if you want to run in network applications as a service um while programmable date, plan virtualization allows us to run multiple intel, um in-network applications simultaneously on a single network device. The code synthesis of multiple tenants on a programmable switch has some additional requirements.

I

In particular, mechanisms for security, performance and resource installation are needed to compartmentalize the network device, while multi-tenancy allows multiple users to share the same physical data plan. There are several aspects we need to consider in terms of security when reusing memory, for example, for packet gaming. We need to ensure that one user cannot infer the data from the traffic of other users.

I

We could allocate a part of the memory for each user or make sure that users cannot reference memory that has not been erased, since all network packets might share the same physical pipeline, one user could impact the performance of other users on the switch by continuously cloning and recirculating packets.

I

Therefore, we need mechanisms that enable a network operator to limit the capabilities available to user applications. Resources such as tables registers and counters allocated to one user should not be accessible to other users on the switch.

I

Therefore we need to and we need to ensure that applications are modular and self-contained. In addition, when a user is deploying a new program on the programmable switch, the operational data plane should not be interrupted from the perspective. Other users.

I

I

Consider an example where alice and both communicate through a programmable switch, alisk has deployed a network function that implements congestion control by looking at the queue occupancy.

I

Where the queue size is carried over, the metadata bus is an attacker that has been has deployed a malicious malicious program and he could attempt to modify the queue size on the on the metadata bus um by setting it to zero or maximum, which could lead to incorrect congestion control.

I

It could also um deploy a program that could attempt to change the header information, such as priority and attempt in an attempt to cause a packet to be dropped to prevent this type of attacks, and a network operator should be able to control what capabilities are available to user programs.

I

The operating model of multita and data plan should allow a network operator to set actions before and after user program.

I

This approach allows not only to programmatically configure the assignment of packets to each program, but also to limit the amount of outcome track. Traffic protocols such as vxlan can be used to encapsulate packets to enable users to modify inner packet headers, while the outer headers are used to forward the packet to the correct destination device. This method can be used to abstract the underlying hardware infrastructure and provide the ability to users to perform arbitrary packet processing, for example, to modify ethernet headers.

I

To manage multi-tenant data plane, a network administrator can can apply rows and permissions to define the level of access available to each program. A super user role can be used to define how packets are processed before and after user programs, while a user role represents all capabilities available to user programs.

I

In addition, a network administrator can define custom roles with specific specific set of permissions. For example, customer roles can be load, balancer and vpn. Routing all the permission to recirculate packets might be necessary for vpn routing. It may not be required for load balancers.

I

To address these requirements, we propose mtpsa multitap programmable switch architecture; it extends the portable switch architecture with the ability to support multiple users that provide a super user part and provides supervisor pipeline.

I

The supervisor python enables processing, before and after user program, it can be used to programmatically associate incoming packets with user programs and to apply packet, encapsulation or other operations on the incoming and outgoing traffic, such as limiting the input and output ports available to a user program.

I

um We implemented the prototype of mtpsa on the psa bmw software switch using the p46 firework and, as you can see, a supervisor program is used to define the ingress and egress pipelines, while user programs are similar to the simple sim switch architecture, they consist of programmable parser, match action, control block and a d parser.

I

The super user ingress pipeline assigns user identifier to each packet. It corresponds to switch context where the user program is installed by default. The user identifier is set to 0 and the context corresponding to this is reserved as a control context. Control context is used to process unexpected packets, for example, when no programs have been installed or the received packet cannot be associated with any user.

I

The ingress pipeline can also be used to programmatically set permissions to each packet using the standard metadata bus. The mtpc target is using the value divided to limit the capabilities available to the program that is processing the packets.

I

The ingress pipeline can also apply packet encapsulation by adding outer headers to the packet. This approach allows us to provide information such as user identifier to other devices within the network.

I

The egress parser is applied to packets before the user pipeline and enables packet encapsulation transparent to user program by extracting the outer headers of the packet, the either smart action tables and d parser are applied after the user pipeline and can be used to process the outgoing packet before it is being sent out to the output views.

I

This method allows us to remove the encapsulation headers before sending the packet to destination server or virtual machine um to evaluate the hardware visibility of mtpsa. We implemented the prototype on the net pjs platform using the 7xp4 estimate compiler.

I

This prototype allows us to measure the packet latency throughput and the amount of allocated hardware resources, and it currently supports eight user programs and it can be extended for simplicity.

I

The user programs are installed between ingress and egress pipelines and we use 128 bit metadata bus from which only 40 bits are made available to user problems.

I

Our evaluation results show that mtps say scales with number of users and allows users to define programs with different complexity without affecting the packet latency of other users.

I

In summary, we propose a methodology for multi-tenancy and programmable data plan data, plane, virtualization, providing performance, security and resource installation capable of processing buckets outline rate. Our implementation is open, source and available on github. um Thank you for your attention and happy to answer any questions.

A

ah Thank you, rajasthan. um Is there a question, um I think, uh for the people on on this group who are not too familiar with with b4? Actually, we've had a a really wide range of topics today, which is good um for people who are not too familiar for with p4.

A

uh Can you tell why uh multi-tenancy has not that much been used in p4, especially if you consider the problems that are related to uh security encryption? All of these things.

I

um I think it's because it's still an early project, um the language itself has um the p416 version, which was um which is still in development and the portable switch architecture. Psa is also a new proposal. um The specification itself and the new work that is coming up, the portable nic architecture is also very interesting. um I think there are a few papers about discussing how virtually virtualization can be supported, and this is basically divided into three three categories, so the first category is uh power level.

I

um This is very common, actually, um basically to be able to merge multiple programs um together and then deploy them on a normal switch like tofino switch.

I

The second category is a hypervisor based where there is a specific p4 program that can basically do everything, but it can be configured through the match action tables. So um basically, this this approach is using a special compiler that transforms the p4 program into configuration that can be deployed at runtime and the last one is basically when we have an architecture that supports um multi-tenancy.

I

The difference is um basically the first two approaches: don't actually provide performance installation. So if you merge a large program with the small one, um the latency will be the same for both of them um that the second approach has a downside that it consumes a lot of resources, even though um not everything is being used and the architecture approach is still something new.

I

um Does this answer your question.

A

Yeah I'm on mute because it seems somebody's doing some repairs upstairs and now I don't hear myself. um Are there any other questions.

A

Okay, um we still have a few minutes. Do you you hear that noise behind? Well? If they stop there, I think they're sanding, the floors, no worries uh whatever it's a nice background information. um I can take just a few seconds to talk about the uh the nsf uh workshop. um I can share some. I can share at least the picture that we had.

A

So um please please stop sanding the floors, um so essentially, I just wanted to tell this group that we had uh with again henning schulzrin of columbia university. These um um three workshops. In november, um we went across the u.s elections, which was interesting uh where we looked at the future of research in broadband and nsf. Does that uh every four years uh to see where the we're, through four or five years, to see what is the status of research?

A

uh In broadband? uh We had three workshops: one on technology, one on economics and one on digital inclusion and, of course, for this group. It's mostly the technology part that is uh important.

A

um I have to say that the report is being published. It's going to be public, so when it is issued, I will make sure to post the the link to this group.

A

um So I think in the technology world, if I can summarize um by maybe two words, it was data and metrics and both of them are related to a little bit of what being said today and is going is being addressed in this group.

A

um There was a recognition that you know there's not that much broadband research as there was before and a lot of it has been uh allotted to like optimizing data centers and making sure that you know cloud infrastructure works well, but that there could be other uh interesting approaches as we go into more uh in artificial intelligence and networks and as also broadband is more and more wireless and not uh and not wired.

A

We had in our group of panelists, which were invited the people from princeton uh the group from jennifer redford rexford, who are doing a lot of in-network computing. So there was a little bit of a, I would say, a um a discussion about the use of in-network computing.

A

There was also the idea that there's a lot of new satellite networks that are going up and these satellite networks will be connected also to data collection, data distribution and computing centers, so that was actually also part of the discussion and who talks data talks also metrics uh the metrics needed to actually direct the data acquisition and the metrics that are used to evaluate the data.

A

So there was strangely enough not a lot of discussion and, I would say, fundamental um network research in terms of what has happened in the fa in the past, like you know, in terms of more advanced antennas and stuff like that, but there was a lot about uh again how broadband research and network research in particular is moving to be very much data oriented and, to a lesser extent, but also related to uh the topics that were addressed uh by our colleagues in economics and in digital inclusion.

A

uh This aspect of broadband, not now just just not being having a fiber industry, but how people can actually use that fiber, which of course, uh is being now experienced by everybody on this call that we're working from home and I have a lot of broadband, but I have standing upstairs. So how do you deal with all the socioeconomic and geographical relations to access to broadband?

A

I don't want to do more than that, because we're running out of time we're going to meet at the ietf.

A

A

Meeting is in the last session of the friday, which I think is afternoon for me- is probably middle of the night in asia and in in the rest of the world. uh Probably california. This time is in a good way, is in the good shape, uh we're going to have the same type of format that we had today, which is uh research papers or research updates, and if I have a better update on the nsf I'm going to, I think the the report should be ready by then uh and again.

A

If you have uh presentations, we've already had uh two people asking for presentation and we're we're obviously asking for uh for more again we're not going to concentrate on the drafts. We would love, however, to remind some people on this call that your drafts are expired, and so please uh decide what you want to do with them and um move again the discussions on uh on drafts to the list.

A

uh This was the or like some kind of our first trial at doing this more into like a um technical kneading of some sorts and again, I think, we'll have to uh sort out how we want to continue that model. But I think it's it's. What we decided would be the most research in the irtf and a little bit less a bunch of drops presentations which we've been very lucky, because we have a lot of them. But again, this is for research and- and we want to really focus on that.

A

Please please, I was going to say even and jeff, please add I would you.

D

What I wanted to say was that um you know we recognize that there's probably some balance um between all drafts or all research proposals, and so ideas around how to um organize would be very well um appreciated.

D

I in particular like sitting here you know, as we um ask each of the um presenters you know like how does this revert or relate back to coin rg, or maybe you know, having dedicated uh sessions to a particular topic or part of the architectural space, or um you know focusing in on the different flavors of compute in the network. You know organizing principles like that, so we're absolutely open to um you know best practices or requests that you might have um to help us.

D

uh You know balance the the dynamic in the group towards the pragmatics of drafts versus the. How do we progress the area and really work towards our objective to organize the research group um going forward or whether that organization means scoping it differently um or directing it in some uh new way? So your ideas are very much appreciated and solicited.

A

um Jeff, you want to add something.

H

Yeah, I think we can discuss after this meeting and also welcome the participants in the group can feedback to us how we organize better.

A

Yeah um again, I think we're we're we're finding our our footing um and uh I think yeah we we we tried to uh take into account uh a lot of the comments we've had before, but obviously new comments are also welcome, so uh we're going to finish on time, which is really great. I would like to thank uh all the presenters uh alex uh leon and paul um edgar and radistan uh for their efforts and helping us put this together.

A

um We will of course, uh produce some minutes when they're finished doing the sanding upstairs and um oh, my god, this is horrible and uh we will uh hopefully see you virtually on uh march 12th. Thank you so very much.

H

Thank you and meet you in itf next itf meeting.

D

It wasn't so bad the standing, oh yeah, it's because you're not you're, not in it. I know I know to those on the other end. I can assure I mean just to tell you.

H

It was worse, it was worse for you than it was for for us. I.

D

A

um Some stores now everything was closed for since december 24th and I think I'm just going to uh to go. Do some shopping while they're doing that, because you know I won't be able to work here. That's terrible!

A

Thank you so much everyone uh and uh yeah. Hopefully I figured out uh colin and even company. You know that it's been a year since we actually met in person.

I

uh I guess it must be here: yeah.

A

It's crazy, isn't it.

H

Yeah, the next one's not going to happen. Is it something.

D

No, it's not san francisco.

H

A

Phillip, I'm sorry to have to tell you this, but the british, the british well colin. You were in the british trials too uh the british uh version of this being now propagating through north america.

A

um I'm not very confident about the fall one either.

A

I I have to plan a move from my apartment in boston to to here, and I at this point I wouldn't it's gonna cost me hundreds, if not thousands, of dollars to take tests and hear.

I

A

From one place to another and I'm waiting because anyway, I have until june 1st to do that, but I have an impression at the end, the only thing I'll be able to do is to send a company in my apartment in boston, they're, going to pack everything and send it here, because I don't know when I'm going to be able to go.

A

It's just completely. It's just completely crazy.

I

Yeah, it's not good.

D

So colin, we would especially um uh appreciate your wisdom about um you know uh how to take the this idea of um research discussion and maybe even more pointed um research discussion where we cluster some of the topics yeah with different meetings.

A

Having themed uh themed meeting could be good, I think today we were just about four teams yeah. I thought.

I

It was fascinating, each.

A

One of them could be a different team, but yeah. We, I think it's a good. It would be a really good idea and.

D

You know, maybe there's also a small set of questions that we ask our speakers.