Internet Engineering Task Force 101, 22 Mar 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: IETF101-ROLL-20180322-0930

Description

ROLL meeting session at IETF101
2018/03/22 0930

https://datatracker.ietf.org/meeting/101/proceedings/

A

Hi good morning, this is the wrong session. Welcome to everyone. We want to start the meeting.

A

Please be aware of the knot. Well, you know we are not going to read it, but you should be aware that all decisions are aligned with this.

A

Please accept our meeting materials to track the medium. This is the agenda for today, I.

B

A

Keep on and take the name from the ITF webpage.

A

A

Thank you. This is the Center for today to have until 12:00. So it's going to be funny. Some comments about agenda.

C

Georges georges its artists, artists, its my colleague, whole prison,.

A

Teddy okay, we have update the milestone list to more realistic dates, so we are going to try to keep this up to the rate.

A

Yeah send the link that was provided by the ITF.

A

Okay: okay: this is the status of active internet draft. Some of the drafts are imposed because we are waiting for people to be interested or not, and work on that, like young models.

D

A

Some of them I want to be present today.

A

Okay, we are going to have a presentation on spur a little with the sixties. That Michael is going to present him role enrollment priority, but is using sixties and okay. We sent an email about the IPS that we have in the current documents. So you have some comments on that. Please into the mini lists are now ice. I.

B

Wanted to make I wanted to make a few comments. It's the first role in our room chairmanship, which takes so long. So it's a bit of an experiment and I hope it will turn out correctly, but we have lots of new new subjects and I think they're worse. Well, the discussion, especially the discussion about how to use beer, is in role on how we're going to forward that. So we all think that needs some time.

B

On the other hand, we have a few drafts, which has already some time there, and we really want to conclude them this year. So I would be so that it can take on all this new world yep.

E

Relating to load balancing, we have created objective, that is a traffic aware.

F

E

Okay, yeah, okay, yeah I feel like I'm eating it anyway. Yes, okay, so there have been some existing standardization efforts for objective functions. There's the two unknown, which are standard and Qasim, has proposed. Load-Balanced objective function.

G

E

Problem with using the standard objective functions, as described by draft in kissimmee as well, relates to networks becoming unbalanced when you use them, so some nodes will get overloaded when, when you use those objective functions that would typically lead to lower Network and the node life and as a result, you will have higher packet losses due to kill, probably and higher packet delay. Well, this is very.

E

So I have example here of one network in which all nodes send data with the same packet rate and, as you can see, it might be unbalanced.

E

In this case, a balanced version would be like this, where it's half and half now the objective function proposed by Kassim will create will resolve this problem because it uses the channel count to balance the network. However, if the nodes, don't all transfer data with the same rate, the result will not be balanced. So in this case, this node doesn't actually send data with the same rate as the other ones. So note that to balance it, you will actually need to make it like this, so the child node component won't be the same.

E

So this is what we are trying to accomplish.

E

So new metric recorded packet transmission rate and in our case we use the packet transmission rate for each note. So it's a local information, but Derek I don't recognize, but he proposed that we could also use an accumulative version of this, where the packet transmission rate would be accumulated. The cumulative thing for all the subtree.

E

We also use a predefined time unit for the period which we use to transfer the packet transfer rate, but we could also introduce a sorry we measure packets sent per time unit and in TSH context, maybe it doesn't matter, but if you want to be not general, maybe you need to measure octet per time unit.

E

In any case, our objective function basically select the preferred parent by the parent, which has the lowest PBR. So this is an example of a dia which contains the Dagon sea and within the dark matter, container data. You would have something like this: a new routing metric container type, which needs to be assigned and just a two octave packet transmission rate. So this is the current state.

E

We have some very, very preliminary results. We have a network here, we used Kentucky to simulate it, and we have you can't really see it, maybe in this diagram, but in basically the network is heavily lopsided or this part so not to and dates and data much faster than the non vult ones. So this, if you know the balance the whole network by either ripple you, it won't work well, but even with channel count, this mode will typically have more children, although it should be more.

E

Some of these heavy note should be assigned to four and the others to five. So it's a bit skewed, so we did some experiments with that and we compared how the traffic will be balanced with load balancing.

E

So we have some improvements in PDR.

E

The load balancing network already gives better results. Ours gives a bit better and mr. Huff tends to slightly worse due to the load balancing issue. We also count parent changes, so you will see there's a lot of variability here. That's a side effect of the very preliminary result, but in this case as well, our objective function is a bit more stable and deletes to fewer parent changes, and we also can't from any di o--'s who are so.

E

The idea here as well is that the number of Dao sent is an indication of how stable the network is.

E

So in this case, we do one and we have some issues already so after discussing the whole idea, we that we need to discuss how general we want our solution to be so a lot of things that moment are predefined, so we you have gotten to a constant one idea would be to either keep it as it is, or also keep a few bits in the metric to send also the data unit, so packets are octet or whatever. Another issue is the time unit, and that might also be a good idea to send it.

E

So you can express a wide range of very values for time period. So the are you talking about packets per hour, packets per minute packets per day, whatever, and one other issue is that we might assume that all nodes have the same forwarding capacity, so that might not be very reasonable.

G

E

We could normalize the PTR to a given capacity or you could send the max capacity as another field as well. So obviously, this has some network overhead as well, and we also see that all nodes consume energy with the same rate. So you already have the energy metal container, which you count energy remaining until the end of the battery, but I haven't seen something that actually counts the relative consumption. So we can deduce that from the to some extent the how much energy is remain with you. It might be more accurate. We use and replace.

E

Finally, at the moment, the packet transmission rate- again it's it's derived from what we note. What a note is forwarding in terms of traffic and it might be a good idea to because at the end, what you care about is the whole path to the root. So maybe that would be more interesting. So do a cumulative version of this metric.

E

So at this point we will have this discussion yesterday. Some of these metrics might be interesting to track them to the a B information element, so that notes are better able to do the whole join process when the network obviously can be rolled into one of the priority values we discussed yesterday, instead of being a separate thing.

E

There might be things we might be missing and I'd like to find very helpful comments. I think. That's it.

A

Thank you coming so.

C

I think our discussions yesterday, we had with Pascal I, can also teach guys I think. Maybe this is not interesting to be part of the EB pocket for the moment, since it is layer, three information, that's what you think.

G

So I definitely think this work is interested. I just want to inform that we have done some similar work and I'll be able to get some data. Ok,.

E

Maybe you should discuss and see if we can find some common ground.

G

We said that we have not experimented but six dish. We had sex with single channel mode of operation. Ok, great.

H

Thanks Michael Richardson, can you go back to your diagram.

E

H

One, the one with the seven and eight sorta, the one, with node, seven and eight and two or something yeah that one so as I understand it. It would be desirable for node Inc to pick five as its parent, because no two has forests yeah. Let's do it. So these.

E

Two, let's say: seven and then a conversed with five, some coming back so.

H

That's what I was actually interested in was was node seven. If you want to also try to balance the children, then it seems that you very much want to make sure that seven connected to node 4 such that note 5 has the capacity to accommodate no D and I'm wondering how do you do you kick seven over that way because it's it's a it's not sending traffic really, but it is occupying yeah. Oh no,.

E

So so that's my mistake. All nodes, all of these nodes are actually sending data. I went through the details of the simulation all.

H

Of them are sending yeah.

E

Yeah, so these are sending one packet per second and the rest are sending well one point: one packet per I, where I is the idea? Second, so 1, /, 7, 1, /, 6, etc. Ok,.

H

But what I understand is is I can see how you can, how to in an eight, you know, need to make intelligent choices and about there I'm trying to understand how in your algorithm, does node seven, which is not a high sender. How is it discouraged from connecting to five such that number? Five has the capacity the child, not that the bandwidth compared.

G

H

Slot capacity, because if five was overloaded in terms of children yeah, then it would be unable to accept eight because but it just but seven could go the other way right. So this so.

E

This goes into the details of the simulation, even if I remember correctly, the default settings for the simulation allows up to eight neighbors. Okay, so I would guess that wouldn't necessarily be an issue in this example, but it.

H

Might be good to do some experiments where you have enough nodes such that you actually overload the child capacity and then see if you are still able to migrate or partition your network, given that you have children that, like 12, has no choice, it has to be the other one got no choice, they have to be there, but 7 is critically. It has a choice. Yes right, so these nodes are basically here just to add some.

E

Traffic to five, of course, but you're right yeah, so.

G

When, when voting in this context, just because you know so, the objective function here takes into consideration only the packet transmission treatment in yes Union. It has to be clogged with some other matrix as well, because packet transmission break alone. If you end up using all.

E

Right so normally what we would do is use some other metric as well like ETA, X or something else and use the packet transmission as a secondary metric. That's.

G

One actually things between so.

E

There are some problems with that in context. Specifically, it doesn't allow multiple metric containers, so that was a bit tricky but yeah. We have that in mind as well. Another.

G

Thing you have already mentioned is using cumulative metric, which is which is actually very key. It's an important thing. Yeah.

D

E

You thank you. Yes,.

D

I'm Charlie Perkins from future Huawei I, said part of the simulation parameters that all the nodes have essentially constant transmission right.

E

Okay, so the period is constant, but the difference between the nodes.

D

Pretty wide like widely applicable thing, I, but those who have varying flows and then you might be subject to some sort of oscillation. So.

E

We're talking about the press, the traffic right, yeah yeah, you might have that that's a problem with all the objective functions. Whenever you have like some up the changes you need to be able to adapt and, depending on the video, how quickly you, how often we get the D iOS, they update, might take some change and you might have some problems, yeah for sure.

D

How stable the solution is given various the SE of parameters for variations.

D

Thank you, yeah.

E

So for sure, we're going to do some more extensive experiments for that Thanks.

A

Thank you Aires, who has read the drop, who is willing to review the draft grain? Thank you thank.

E

A

Is present in the next so.

E

This draft we have presented it again in Singapore. We have some changes, probably go over them a bit copy, so we are creating a new TLV in the NSA in the note state and attribute object within the metric container. This is used basically for enabling PR a packet, replication and elimination. So, since the last version we did some editorial changes. We improved some of the diagrams and some working SR model. We specified how some of the vacuum see, fields should be used and we also created somewhere shorter dissectors for this for this field.

E

So the here is that we were trying to as much as possible achieve some determinism to get some reliable communication and some low jitter performance as part of that. So the idea is to use packet replication illumination. So you replicate the packets to multiple parents. You make sure that whenever they arrive at a common node they get eliminated. So you don't get a storm-off packet and they also use promiscuous overhearing to increase the number of packets they received.

E

So it's more extensively dispersed so in this draft, we're basically going to work on the alternative, parent selection. So, in order to do the replication need to choose where to replicate a packet to, and this work relates to the parent selection to whom to replicate.

E

So we need to extend the DAO messages with some information to allow child nodes to select an alternative parent as best as possible for our purposes during that, for that parent selection, our draft enables it to function at all, and specifically, the idea is to allow selecting an alternative parent which has a common ancestor with the child parent. So, in this diagram the idea is to be able to select B, which has a common ancestor with a so then B needs to be selected for that purpose.

E

Also, very quick example afterwards, so we have an again of the DAO packet with the dark matter container and in it we have the MSA object and within the NSA object will have a new request for a new TLV, an optional TLV, which describes the information which has information used for PRA, and this information is basically a number of addresses of ipv6 addresses for the parents of a node. Now.

E

Okay, some of the new stuff is how we use the MC fields when it was requested that we fully specify them. So basically, we define that it's used as a constraint with had allowed it to be unspecified before and yeah.

E

That's basically, that's mainly the only thing that we needed to define that if this is specifically a constraint and not a metric, the other fields are not already affected and we also implemented a washer dissector for it too, to make sure that we are correctly creating the field, and so here is a quick example of how this would work.

E

If you have this network- and you have, this is the preferred parent of s and D is the preferred parent of a, and this is the parent set of s, and this is the parent set of a and you have some intersection here. This is a parent set of me. You want to select the sorry, the the B node as the alternative parent for s, since there is an intersection here. So since B is in the is in the parent set of both a and B, you select B as an alternative parent.

E

The idea here is that whenever you do the packet replication, the replicated packets won't veer off too much from the original path and the reason for that is to keep jitter and the latest in control.

E

Now one issue you have is that, unfortunately, sending all that data is pretty heavyweight like 16 bytes per ipv6 address. One thing that we have done is to remove half of that to somewhat compress it, but again, that's not ideal, and we also have some ideas. We we are thinking about using this, the same information not for keeping the route of the replicas close to the original path, but intentionally moving away from the original path to get some extra diversity, both for just the preferred parent.

E

So even without alternative parent and also for a multi path. We want to do a replication, so this is it I, don't know if this is interesting and if there are some leaders to continue this work, I had hoped to present some extra experiments, but well.

H

Michael Richardson to compress your v6 addresses code. That's probably already present would be in the sixth orh header, which basically has a list of our addresses and comprends since they're similar. It does a very good job of compressing them into maybe one or two bytes, each. That would be ideal and the codes already probably present on most systems, so I'll.

E

Keep that in mind I'll take a look thanks.

I

Justin I have a non-technical comment. You have a draft with an object that it's called NS a yes.

E

Unfortunately,.

I

E

I

E

I

E

Always saying pns, but it never occurred to me to.

B

Stock and bit curious I mean we just wrote you about parent selection, and there was an algorithm and also other people who have done graphs about parent selection. How do you think this proposal will actually coexist with those parents, election drafts? Okay,.

E

So that's obviously an issue depending on what you want to achieve. I don't know if you can necessarily do both at the same time, but so what we use at the moment is use a normal metric like ETA, X or whatever, and use our constraint just to constrain the alternative parent. So it it can. It can work with other metrics, pretty okay, that's not a problem in itself.

E

I hope to present some result is when we have it almost fully implemented. I hope that I will finish that petition yeah.

J

The objective functions that usually we present at the ITF are very simple, like use one metric, but we expose a lot more metrics and now we put more and more stuff and real world objective function may be more intelligent than that. It's an objective function can be a piece of logic. It's meant to be a piece of logic which ties a number of metrics together and the way it's tied together, maybe Oh. First, you look at this because that's my most important concerns and the other things maybe tiebreaker something in these guys.

J

You really want to build two non congruent paths with the capability to do, but back replication and elimination in the middle of your network, because network are so lousy and you really want your data to go through in the definite definite time. So that's your priority. Now, if you can do that, and if you see multiple parents which would fit then all of a sudden, your whole life may be taking more ideas from some of this component works that we have used up the Essbase, TTX or use the load.

J

Balancing that we just saw I mean that's. How it's it's rebound together? It's the fact that your objective function will not be one of those. You know simple apps that the IDF has produced, but one which really does what you want. So if you have an industrial consortium which really cares about deterministic properties of closer to deterministic properties in this network, you will want this and you will try to do some traffic balancing if that's possible.

J

If I have multiple solutions, then, as a secondary, saying, I will look at the load or I will look at the number of children or whatever house right. So that's how it's bound together.

B

Would be nice if you graph could show that.

A

Focus right, the drop okay voice will interview it, and this work is important for 60s right.

E

Well, parts of it are so this draft is specifically, you know, but there are parts like support for overhearing, which we actually need to create a new cell type in in the schedule for six days earlier.

A

Okay, thank you.

J

Bicycle again: well, sis teach does that need that right, but a solution that would really benefit from that would also really benefit from 6-2 to make those two work really well together, you need a six infrastructure, so we can schedule stuff and yeah, and you need this because after hearing really means that you need to know where and where a certain stoker will talk, so it will able talk to B, but C will be listening, and so, if you have a some form schedule which allows people to know when this transmission occur along this shape that we have just formed, then that's when you get the most benefit of it.

J

So six ish will not say we need that not at all, but the solution which needs that will also need six. Okay. Thank you. Okay,.

A

um We screens reviewed sorry again.

A

Okay, thank you. Thank you.

G

Hello Java from Huawei technologists, so I am going to present ripple observations. So basically, these are just observations, no solutions here. These are some of the issues that we found during our solutions: analysis, design, deployment, implementation deployment pilot. Actually, so we we ended up having some sort of solutions to some of these problems, but we definitely believe that those solutions are not optimal, definitely not best in all the cases, and we wanted to bring these problems to working group to check whether these make sense and if we are not missing a major point here.

G

So that is what we wanted to check you. So our our deployment, our pilot, was primarily towards smart meter networks where we have thousand know the cardinality with sixteen hops, which is quite big, and we have storing, as well as non storing mode of operations involved. Having said that, most of the problems that are listed here are related to storing mode of, but we definitely believe that some of these problems can be solved in a better way. Okay, so first problem.

G

This is all of the major problem that we have faced is is how to handle the DTS in increment. This is non-trivial, especially in storing mode actually non, storing mode of operation. This is pretty easy to take care of in storing mode. How do you increment the DTS L, so the the decisions that you make that can impact the downstream route availability?

G

It directly impacts the control traffic overhead for the DAO and, overall, it has a major implication so on overall number of pattern, switchings that are present in the network- and we think that you know this has to be handled and not not necessarily by I, don't know the the suggestions could be completely different could be completely so one thing I wanted to check out is how beer or some other proposals, how can impact how those proposals can impact this these problems, so that is still under consideration that we have not read those we have not still considered or understood those drafts clearly to check whether those have some implications of this problems, so there is so depending upon.

G

So this is the first implementers deployer dilemma that we we had so, in which case should the DTS and B increment? So DTS n is the number which is it's a sequence number, which is part of the DAO message, which essentially is tells you which tells now the child nodes in what errors, whether it should send the DAO message or not. Now there are two problems here now. There is currently no way for the target node to know whether the DA was actually reached.

G

The border router, the end-to-end path, has been established, so it needs to so the current network. The current mechanism has to have enough redundancy now redundancy so as to make sure that the DAO actually reaches the border router and after that it could. It should ideally start the application traffic. It is very important because if you end up starting the application traffic before the routes are established end-to-end, then you end up clogging the network. You end up queuing.

G

The acts are not going to come back and you don't want to ideally start your application traffic this time. So essentially, what so? Should the DTN be incremented with every DI electrical timer now I know for sure that at least the old implementation of quantity incremented the D TSN in every trickle timer time out, which is bad. But then there is no option. I feel because the moment you have a higher number of might the number of hops increase in case of storing mode of operation.

G

You have to depend upon this particular increment thing of DTS n, so that you ensure that you're now actually reaches with some confidence to the word router.

G

Now, if you don't increment DTS in in the electrical timer, then your doubt redundancy is too low your you have a high probability of now not reaching the border router and, of course, with the increase in the number of hops. The probability of now success drops sharply. So this is the something that we had seen happening in our networks and what we had seen is actually the network convergence time 90% of the nodes gets joined, but the remaining 10% of the nodes.

G

They do not join for a very long time and it's very difficult for us for us to actually ascertain whether you know so. We found that this is one of the primary reasons.

G

Second, in case of parents which should the node the increments it's DTS in well, it looks natural to increment the DTS in because you want the sub child's, not only the sub charge, but the subdue dag routed at that particular parent to updates its routing.

G

But the moment you do that you have a lot of drug traffic because the whole show are subdued, so you go on incrementing, the DI or DTS n, not only you, but your sub channels have to increment the DTS N or the complete subdued air has to go into the I mean it floods, the network for that particular subunit.

G

So that is what we have seen, and so is there any better to operate? This storing mode is the mode of operation that is considered here in non storing mode of operation. Again, this is not a problem now. This again is one important point and it has so Dow AK. This has some relevance to the previous point discussed. The DT is an increment. If you end up handling the Dow back properly, then it will solve good number of issues. That is what my opinion is.

G

There had been discussion on the mailing list in 2015 regarding this, and there are implementations, for example, with the current ripple spec. There are multiple interpretations possible whether the X should be sent hop-by-hop or it should be sent end to end, and there is no way that this simple I mean once currently riot implements hop by hop acknowledgment acknowledgment. The older version of quantity implemented hop by hop acknowledgment, but recently quanta. He found that we have they. They they need in to end acknowledgement mechanism, because the primary reason is the target node. Without this mechanism.

G

The target node won't be aware that it has reached the border okay. So how do you should then implement or interpret this particular this particular scheme in in the ripple specification? There are pros and cons for each of this mechanism. Hop by hop acknowledgement is actually pretty easy to implement. There is no state involved in case of end-to-end acknowledgment either there is some sort of state involved in the routing table or there is some sort of overhead involved in the network controller and network messaging, so it has its own pros and cons.

G

We eventually ended up implementing this into an acknowledgment one in it, but in a different way, without that we couldn't get our convergence time in any good shape. So the other problem here is how do we handle aggregated targets in case of dower, so the our work is for a Down message and not for individual targets, and now message can actually contain multiple targets. So how do you acknowledgment? How do you acknowledge a particular die which has multiple targets and once a subset of the targets fail? This is a problem.

G

They're, a ripple specification says that acknowledge acknowledgement cannot carry any new option, any options, not new options in your existing options. So there's pretty much no schoo for improvising here. Unless and until you have a non-standard variance of implementation.

G

If so, another problem is ripple is not clear on how to handle our aggregated targets. It certainly allows it. There is thus definite wording in the specification which allows aggregated target, but it does not have failure handling for it. This has been discussed on the mailing list in 2015 as well, and I feel this is important to be handled.

G

If you see the two most popular implementation, koa riot and quantity, does it in a different way today it is impossible to get an any sort of interoperability possible between these two implementations at multiple hops. What we have seen is right sends aggregated targets, quantitate doesn't handle it, the network doesn't never gets formed, so the way I have been experimented with is I have a border router and a few of the notes which are counting here, another ad, but this will have a problem only at multiple hops.

G

If you have a smaller network, you are just doing an interrupt at a very small size on a table with all the nodes speaking to border route, but directly you are not going to have any problems. The moment you try to scale it try to achieve performance at a bigger scale. That's when all these problems will start creeping in so now AK is important. I've already told talked about it. You know, because it is important to know when the end-to-end path is established.

G

It is important for the application to know when it should start its traffic in case of hop-by-hop. The Dower can fail or buy up. Acknowledgement is actually not much useful in my opinion, because you have link layer acknowledgement as well, so in most of the cases most of the implementation, disable it by d4. So I don't know if anyone uses hop by hop acknowledgement mechanism to achieve anything specific right now.

G

So there is one more alternate behavior that we have implemented. Basically, what we have done is the DAO goes like this, and the border router eventually sends or the root route eventually sends an acknowledgement directly using a global using the group elect because the downstream path is already established. So this way you can have least overhead in terms of control traffic. You don't have to have any additional routing state on 6l, arse, and but the problem here is you can't handle aggregated acts.

G

So there is still a downside to this particular solution, but we eventually ended up implementing this this this solution and we got a good improvement in the convergence time. This is the mailing list, which was actually it was actually discussed in a more substantial way.

G

The lot of discussion that has happened already, the next point I'm going to discuss I'm, not sure I, would really like to have some feedback here, and this is something that we have faced so in case of storing mode of operation ripple has some state information to be maintained across node reboots in case of I/o devices. The flash endurance is a big problem.

G

If you end up having your network protocol flashing, something to the persistent storage every now and then it's not acceptable. At least this is what our solutions requirement team told us. It's not acceptable, so you have to handle it in some other way, so I'm not sure. If anyone has handled this particular problem, current implementations, open source implementations, don't handle it. They expect that there is some persistent storage, but they don't handle it by itself and we feel again. This is important again. This is more impacting, especially in case of storing mode of operation.

G

H

Coaches are, did they they say so. Do they say that you can't write anything to the flash all relating to the network, or did they say that they there's some upper limit or yeah.

G

So there is an upper limit. Actually that's a good question. Thank you Michael. So if you see the draft I've, given some good numbers in there, I've actually put in the numbers. Okay,.

C

G

If you end up incrementing DTS annual five times a day, it's a problem. Okay,.

H

Five times a day is a problem: okay, one time.

G

A day, one time a day should be okay, but still what happens is now. Your network lifetime is not dependent upon your battery, but the flash endurance of.

H

Course, efforts, yes, um um but of course it now you still have Network battery to tell you to tell the system that that you're dying, because your Flash is already at least which is better than the other way right, yeah but I.

H

The reason I'm asking is it is because there's other things like a SNS and in six station other stuff and network keys that that you do want to write to, but you don't need to write even daily to the system's, but there are some other things that the network stack needs to keep track of. I. Think so that's why I wanted to know if there was some threshold, that is a pain for them and then because your other slides were how often do we increment the DTS ed well.

H

Clearly, the answer is not more than once a day right.

K

H

I had some other questions, but I wanted to wait to the end I think first, because he seemed to have a lot of issues so so.

G

Just to just to mention something in this context, we have done some real experiments and the number is substantially bigger than five buddy in story mode of operation. You know, and that too, with the modest-sized Network first.

I

I

G

Sorry so that you're talking about right amplification where you change the location of the flash so that it gets leveled across multiple listing, so we I have put some numbers in the draft. Please check you know if those make sense. Okay, I really want those to be reviewed and I would really so I've put some specific information, considering the write amplification, considering their node endurance level, considering MLC TLC flashes and then provided some data there.

G

So most of the most of the IOT devices implements CC 2, 5, 3, 8, 2, CC, 2, 6 5 0, for example, implements some cheaper flashes and it's a big problem with endurance, 0, 6, sellers and particularly, are more impacted water out. We don't care mostly because yeah. We assume that it's, it has some sort of other mechanisms to handle it okay.

G

So this is again a hot topic. It's already been a handle, I think other works in progress, and this is definitely impacting the overall solution. Implementation for us basically handling resource and ability in terms of neighbor cash table and routing table. How do you handle what happens with the neighbor cash entry? Goes. Full there is no signaling, currently I.

G

Think part of the work is being done by in the enrollment as part of the enrollment draft, but the certain scenarios which I still feel may not be the draft, might not be still able to handle it. For example, in in case of this, let's say, for example, therefore routing table entries on n1 and n2. Eventually, there is only a single like 6lr here so or rather three routing table entries. So how do you handle? How do you signal it?

G

So m3 has only one preferred parent so either all these four entries will go through this n1 or n2, and both of them have out think. So. How is this going to be handled in? So there are some scenarios which are really tricky, especially when it in context to handle degree, social mobility- and this has been a major major problem for us. This directly impacts the network convergence time.

G

This directly impacts the packet delivery rate for us and we have an implementation which tried to solve this issue and we have not ended up anywhere as of now, but we have a lot of observations and I feel if there is some, if other people have some other alternate solutions, we'll definitely like to understand and if I mean move, definitely implement. If we someone else, has a solution towards this all right. That's all! Thank you. Precious peace.

H

Michael Richardson again, this is wonderful work I really would like to adopt. This is clearly there's a bunch of updates your points about do AK, Papa, hot versus and and I implemented end and from from just thinking about the problem. I said well has to be, and and or there's no point but but I hadn't occurred to me that the spec was ambiguous about that um I think I just assumed.

H

That was the case so clearly, there's some bugs in the spec and I think you've identified those and I think that that the places where my suggestion is that that I haven't looked at your document, I'm sorry but I, think you should split it into ones that are clearly. um There is a problem here. It doesn't matter whether you recommend and end or or hop-by-hop act.

H

So long as the document outlines those two things that there is a ambiguity as to what it is and then the working group can adopt the document and make a decision.

H

In fact, I think we should adopt the document with you know, a or B and as a working group document, and then we should make a decision as the working group as to which one we're actually going to compete if it to a bunch of those things, DTS n, increment I, think that's a bug in the spec I, don't know what the answer is, but it's clearly a bug in this back. Okay, so I think those are updates.

H

26550 and I think that would be a really good document to do and I think it should be in some sense non-controversial right and then you should have a second document for issues like this one which honestly we've known about in some sense from the very beginning. We don't have an answer for I think ultimately, a kill, storing mode yeah.

J

H

I think, ultimately, that's what happens and and so deep things like DAO projection and stuff like that, gives us back a lot of the features of storing mode, doing in a different way.

H

But I, don't I, don't know, I, don't know how to solve the problem either, and it might just be that in the end that you have a document which top takes 6550 ops, leading storage mode, storing, though in the it as a result of this kind of problem or saying, if you're going to implement it, be aware of these issues through which we have no profit, no solution and so think about something else.

H

So that may be a very small thing, so I think that's the right answer to me: that's where to go for it, but the bugs I would like to see that just let's do the document, let's adopt it, let's get it published by the end of the summer, okay, because I think that's like non I think it should be non-controversial once we have the two questions in front of us at the next meeting. You know a or B, let's hum a or B which one you want.

H

I think it will be very conclusive once the once the details are in front of everyone, though obviously it's B or obviously it's a and I think that's just the way to go. So. Thank you very much for this work. Thank you. Well,.

J

I agree with Michael about thanking you so much for the work. I don't agree with much else. Okay, that's.

G

Great, that's actually.

J

Okay, so I can give you some history on some of those things and I agree on the on that fact that the spec is that complete and how you use those things and part of the reasons is: probably people not didn't necessarily agree on how to use them. When you wrote this back, so we ended up leaving it open, so so the next generation would decide which is what's happening now so I'm very, very, very happy to see this so I love your draft as a prime statement.

J

I would not rush to the solution, because the reasons why those friends are still there is because we could not solve them the first time. So don't rush and publish something in two months that that's where I disagree with Michael, because it's not that easy.

J

The dialog was mailed two replies because you're thinking about radios right repos are watching protocol. We run it on wires and very high speed networks. Actually, so that work is real, you don't have any artwork in the initial design was not end to end I mean if we want to design it end to end our watch. Let's do it it doesn't. We don't even have to call it that watch. We can call it the way we want and we don't have to stick to the dirac format.

J

The deck is read there when you can't leverage your lower lay your thing for doing it and and repo as never of things like that. Well, you say: okay, if we ever, if you have a lower layer thing which, which will do it you, okay, the the idea was if you have a Dao information, you need to pass it to your parrot and if it doesn't work, then you need neck nourishment and you retry and if doesn't work, then you have an attachment. You retry at some point.

J

You need to pass it to death and yes, it might. We have to see the consequences, but we hoped it would work after you, ultimately, retrying to the athletes would end up working. We have to talk about what happens when it does not just discovering that it did not. Doesn't help a lot, because you don't know along the chain what's broken, so we have to figure out a solution.

J

It will take a little bit more than a second to get there, but just understand the end-to-end rod, because we are, the initial design is just local at lachemann. It's not enter well, we want something to solve the end to end. Let's welcome it. It's not there. Ok, all right the dts. There was initially something that you would trigger either. If you find the discrepancy in what you see like you know, this flies in the packets which tell you there is something wrong there.

J

I just want to assert that I have my knowledge of my children is what it is and the idea was either us just your first half children or you go as you said down through the geotag. It could just be a one heart with rush I think the option is Taylor. Initially, that was the option. I just don't remember what ended up in the spec. We met so many revised, but in my mind, was either for your first hot children or for everybody one design behind it was to rebuild the geotag without reforming it.

J

If you change the version number, you will reform that you, like the shape of the topology, will be changed right because people will hear her and selection. Now. Imagine that you do a DTS, M and trick yourself. You skip the same. Do diet, you just repaint it right paint by painting I mean we put the addresses where they are on the other, so that was the difference between the DTS n and the version version reveals everything DTS end just retains the existing structure.

J

So that's what I mean use it when you want, but that's what that was the design and the when you want now is what is not described enough and that's what.

F

The friends like yeah.

J

One thing was I, find the discrepancy looking at the data traffic I'm, getting a packet from somebody and I, don't know that somebody, okay, interesting, do I know on my children right, that's one reason. Another reason is you know, I think these guys well the packet going up when you think should be going down all those things. So there is something wrong. I may just repaint my graph to see us behind me. So these these are the sorts of reason.

J

Also another one would be I want to reparent I lost all my possible parents, and I would like to explore who's behind me, so I avoid them and I try to jump to somebody else with not the idea just want to reassert behind me just reassured that this guy's not the enemy. Let me jump to it and take the risk. You know we can stretch.

J

We can actually go down a little bit, so that's one way to make sure where I'll make sure to hopefully don't, but, but that's not for sure, but the account but I wanted it.

J

For personally was the capability to rebuild the network router intelligently by having a timer, which is longer at each Welsch, getting shorter at each hub as you progress the DTS and the new DTS, and so that the children would advertise, and then the parents, and then the parents and then the parents, each time after our timer you see and so so enable like the repainting like the way.

J

We never actually standardized that, but that was one way I want that not only to refine the dad without changing it, but also to do it like smoothly like a counter wave. You know away from the bottom to the top. So that's what it's for now can we use it? Is it usable, etc? I'm just telling you what it's for yeah, okay and then again agree with everything about the prime statement, things being and clear, I'm just telling you what the tools are for the arc was local one hop yeah.

J

If we're missing something more is design, it don't call it it back, because that's not what yeah.

G

Thank you. Thank you. Michael.

H

Richards, it again Pascal that was wonderful, I'm, so glad it's on tape, so the the the III think that you're in violent agreement with me actually well, the the the point is that if there's in clarity- and you just clarified about four things- okay- that clearly many of us were unclear about okay, so clearly 6550 didn't say it well, it clearly enough. So that's what I'm saying all those things you just clarified.

H

I would like that in a low-hanging fruit update document, okay, and if it turns out, though that, as you said, we need an end to end thou act or equivalent. Okay, then great, that goes in the other document of of things, so so the the the purpose of the low-hanging fruit document that we think we can do quickly is that nothing in it should be controversial yeah, because something now you just clarified a bunch of things which now makes a bunch of things potentially completely non-controversial.

H

If that's what the the clarification ISM, so that's all I want is so that there's no further, you know concern or whatever about what does this mean or what does this do? In the this business with the bottom up, that sounds great I, never conceived or understood that from you know the last time I read 6550, but that's really I did not know that. So that's really important. That's why I'm trying to capture is, is there's a bunch of knowledge that we haven't all come to and we'd.

H

Maybe we'll go back to 6550 and go oh yeah. That's what that paragraph meant. After all, it was never really clear to me when I read it I needed a bigger example to explain it now that I've implemented in Pascal already conceived of it ten years ago. So we already it's gone. So that's what I want. That's really what I want is to capture all that that stuff quickly thanks.

B

B

Thank you very much, thank you for the presentation, but this.

G

Is clear to you what you're going to do so I I think some of these things we can take it on the mailing list. For example, what are the things which are which, which the working group thinks is non contribution can be used.

F

L

As part excellent.

G

B

G

Part, we can see how we can take documents. Ok,.

B

Very good and document vision statement about the problem and the solution: yeah yeah, oh.

J

That was my question: can you clarify these everything is in there must be captures. They be document blah I'm. Just we just want to understand. What's the future, is this document like a prime statement document and and put it like that or do you want to evolve part of the solution in the same document? How do you want to structure the world? The.

G

Current document, at least is purely problem statement. There is never solutions, discuss that. Yes, I.

J

Mean that I, like it like that, because now we have a repository of our prime know.

B

What I was thinking that we don't want problem statements to happy unless informational as she can separate in different documents, and he can motivate by he's looking at the problem and then give the solution. I think it will be perfect document. Yes,.

J

Okay, so do we keep this one and then write new documents? Well, you get a piece of the prime end and the real solution, pointing I mean I'll. Do structure yeah.

A

We can have a maybe main document with the problem statements and then that one as well pointing to the solutions you don't.

J

Have to publish the Prime statement, we just keep it alive, yeah and then yeah. Maybe everything fits somewhere, but if not, we still have this so.

B

Your proposal to heaven draft is to complete a problem statement and then this draft is slowly disappear once all the documents are published. That's great that depends on or you see the work we.

J

Did that at six law right I mean we had a problem statement draw for a long time now it's dead, but but it's it's actually in appendix four, the main spec which solves that.

G

Yeah so basically you're saying that we keep this problem statement document as a as a things to drag.

J

C

Of the sections.

I

Earlier life, when I was sharing the rock working group, we had a standing document for about five years that documented problems.

G

I

Things frequent misunderstandings, and finally, we published this at some point, but we left it as a standing document for a long time. I think.

G

Custom is talking about clause in ll, an experience document which is actually.

I

Know for sure, so that was.

G

Clearly mistaken.

I

G

Corrections we.

A

Have this case where, when the michael right there.these template for applicability statements, it was like a drop and then, when all day, thank you and yeah. Thank you very much.

B

A

Was ready to drop.

A

Quiz will interview it right, yeah, we need it. Thank you.

A

B

We have this discussion now. If you want to introduce beer into ripple, there has been an earlier document by kirsten who says how you could do multi multi cast by using beer. There will be another presentation later on by Pascal to see the different modes of beer you can use and also for unicast and having all these proposals we wanted in the end to start and Design. Group is the agreement which the working group to look at this problem further, how to put beer in to roll a repo? Sorry, yes, Carsten! Please.

I

Yeah, thank you. So yesterday, in every guy was talking about stuff. That was not and here's no. We are talking about things that that is more than five years old, interesting nostalgia based.

I

So basically the the observation was that it would be good to have a multicast phone on story mode and we were thinking about good ways to do sauce routing for Manteca. So that was a research project that we started some time ago and we came up with a solution that we call constraint cast and that actually yeah was a result of a research project, but not necessarily a draft that we wanted to pursue here in the ITF.

I

Then in 2014 beer happened and we saw oh, there are other people who were working on source voltage miracast. So it's not exactly sauce routed in vo necessarily, it's maybe routed somewhere, it's but yeah. So this draft. Suddenly this research suddenly became interesting again and we submitted it as as a internet draft and.

I

Spun for a while and became a working document, but it certainly makes sense to do a revisit of what's actually happening in the working group and see how that is applicable to what we are doing, but maybe also the other direction. How what we are doing actually can be useful in in some deployments that what via currently has doesn't fully address. So I have presented this before just to remind people what this was about. So sauce route, a multicast in non steering mode, the source routing always comes from the route.

I

You need to put information into the packet for every forwarding node that sees the packet to decide whether it has to forward it or not. And of course, one way of doing that is just listing the forwarding nodes in a sauce rod, but that becomes big very quickly, even with with great six large compression and so on. So what we looked at was what what are efficient representations of lists of data and since 1970. Actually we know about bloom filters.

I

I was eight years old at the time and so yeah, that's an technology that was invented for the the machines of 1970, which are constrained nodes by by two days thinking.

I

So, let's pull out that that idea and see how it works, and basically the idea is we put the outgoing interface addresses of the forwarding nodes that are supposed to forward things into the bloom filter, and so every node can decide whether to forward a packet that it sees or not.

I

So we can directly use the occurring interface address. We don't need any coordination numbering whatever, so this is a small incremental layer on ripple non-selling mode and of course the problem is that bloom filters are probabilistic, so we have false positives for its positives, cause various transmission and the spurious transmission, of course eats some energy somewhere.

I

So we have to think about how bad is that in a specific situation and how bad that is, of course, depends on the properties of the network. So there are a lot of things that you're into that. So, for instance, how full are your bloom filters, and that of course depends on how many forwarding nodes do you have for a particular multicast group and also how many bits do you actually allocate for that?

I

So you have a little bit of control over that and if you have a very dense network of course, you will have more nodes that actually see a multicast packet. That is not meant for them and start exercising these false positives. So the density of the network also has an effect on that. So it's it's not something where you just write through the three formulas on the sheet of paper and understand what's going on, but basically the spurious transmission that we are seeing here is not doing any damage, except for wasting energy and spectrum.

I

Of course, so we actually can live with with a significant percentage of false positives. Assuming that the turn ative of putting large saltwater lists into each package are also expensive. They also waste energy in special. So so it's a trade-off there and and the numbers.

I

This is just just an example. So if we allocate just 64 bits to the filter, which is kind of a minimal size, we can have some some six or eight forwarding nodes and I'm talking about for Windows and networking what leaves I'm talking about following out in a magic house group before we start getting 5% for its positives and then this, of course, the handling capacity goes up. If we allocate more bits and if we accept 30% false positives, which significant and then we can do a hundred forwarding nodes in 56 bits.

I

So there is some some scaling possible, but this is certainly not something that you would want to do for a very, very big network, but then numerating bits, maybe also not so great for a very, very big Network. Okay. So that's the basic idea of debt, Draft Pascal! You have to explain what the abbreviation! What means?

I

I

Well, you, you know, I've been the chair of the ROC working group, so I always always was thinking about how to combine rock and roll, but I haven't found a good reason for that. Yeah.

G

I

Okay, so basically the question is: can we do the same kind of sauce voltage, multicast based on on beer, the beer architecture and I'm, not a beer expert, so maybe I should pass kind of the point of, but basically.

L

Hi Kirsten Macklemore, with Itron I, had a question about the spurious transmissions you had on the previous slides. Is that because a hidden, node problems.

F

L

I

The study, so the blue.

I

Dashes offer you and sitting one bit for every hash. So when it looks at it's only bitmap and sees whether all bits on such and what could happen is that bits from different nodes combined and happy to match the bits from from a third node have.

L

You considered the fact that the note on the second half can hear the dag root and virginal transmission.

I

Yes, so that you have to add in some rank information to make the whole thing work, I, just okay,.

L

So your sim in rank is 2 followed. Yes,.

I

Okay, so we are using the repeat structure yet: okay, it's just a little layer. On top all.

L

Right. Thank you.

M

We have any presentation any of these pictures that show kind of.

M

Right so basically, we've got the the beer stuff, which is kind of RFC. Now, um since you month and that's basically, the architecture where you know every bit in the bit string indicates a leaf receiver.

M

Encapsulations MPLS and you know anything else, Ethernet, so that's a flexible encapsulation and then we've just started to adopt by the working group the model where every bit indicates an adjacency. So there would be, for you know the non storing mode where you indicate really the past through the network.

M

There is also, then an encapsulation for that that primarily tries to add a sequence number so that we can do duplicate elimination, press function as they call it in dead net and then yeah the fast reroute things, whereas we're still working around what that should be because primarily I think right now we would be focusing on you know: dual transmission, send the traffic twice through the network and then do the duplicate, elimination through the elimination function and then, as Kyle, had these great ideas on how we can actually use the bits in the bid string on the receiver side to also provide additional OAM functionality in terms of okay.

M

You've been sending this packet two or three times and on the receiver side, you actually want to know okay, which of these packets, which path did it actually go across, and then I can basically do even more diagnostics. With that.

A

Thank You Thomas.

J

First guy here, I've got some backup slides. So if you ask me, because we have time, I can always go through them to show you. Our beer works in with Ripple, with the proposal here. I thought that most people in the room already know that so I just put that as backup slide, but just jump in and say please PLEASE on them, yeah right here or you can visit the slides after the meeting I'll just scroll a little bit and you will find them.

J

So what I wanted to list right here is that, basically, we have this, these metrics of storing mode versus non storing modes and whether you use the classic old beer using bit by bit a bit is one thing versus bloom. Filters were, which can be seen as a a compression form, an alternate form of expressing the bits which, which has pros and cons that now needs to be discussed. But what I'm trying to say first is that the bloom filter can apply without you're, doing storing mode or non storing mode.

J

It doesn't matter whether you expressed the address or whatever else of every harp, which is what's so swatting does and T does so non non. Storing and beauty are really much in the same corner here or whether you want to express the final destinations, which is what normal beer does, which is what our storing mode does. At the end of the day, you can always express these things as IP addresses, and you can always compress those IP addresses into a bloom filter. So it's not a matter of multicast or unicast.

J

It's not a matter of storing or non storing. You can always do that.

J

Disagreement, let's say now: you.

I

Just recorded because moving yes, the guy with the German accent so.

I

So I think it's important to Venus code can compress sets. So it's also normally is the sequence.

I

So there is a little bit of a limit.

G

I

What the bloom filter can actually do for you. So if you set words for you, that's fine and in frost it works because we also have the rank information, so that creates the sequence.

J

So I started destroyed much later than then customs work, because I just showed that some meeting how Bureau could work for repo and then the child asked me to write, write up something so I wrote up something I hoped actually that custard would join. So what I did is I try to walk to write, something which shows that you can do all four and try to present this as a production on top of doing just be available. That's how I structure this document!

J

That's why I have those arrows here which show you that actually, you can do all four, but my main interest when I started presenting this was this this square here and the reason why I'm, mostly interested in that is one of the reasons why storing mode is not so successful in the IOT space is because it consumes resources in individual node and the amount of resources depends on the total amount of leaves that you have in the network.

J

Now, if you, if you express those, leaves as a bitmap so say you have less than 200 leaves in your network a bitmap of 256, give you some elasticity there and you can express everybody. And now the amount of state that you have to keep in every individual node is no more the number of leaves in your network. But it's the number of children that this node can help right, which, which can be controlled, I mean with ripple.

J

We yet don't have those extension which was talked about control, I mean ichiran or what, but that's something that can be controlled. So all of a sudden, even for smaller IOT devices, thanks to beer, storing mode becomes possible so that that was my my lead interest in in using beer for storing mode making, storing mode alive again or something great again.

J

So, basically yeah. The main question that that the group has to sort out is is how we want to structure this I mean. Could we could have a document here and the code occupant sure we could have a document here and and just something which is which there is how you do be a regardless of whether your choice you're talking about is the next stop or whether it's the the end of the path right?

J

We can stretch out this work in the number of fashions. So that's that's. One of the questions we have in front of us and I tried to actually document more that aspect and the actual operation of beer in in repo, but actually it documented that as well, but I kept I kept completely open. How gloom, because I really thought that that would be customs interest.

J

So this is what I'm trying to express here, keep in mind that beauty is all about specifying bits for each hub or each segment that you want to follow. So if, if you see, if you look at what Kirsten Express is in constrain cast, it's it's pretty much. The same thing right, we express the different hubs that the packet has to follow and then the false positive or I computed, a mix of two uh shares in and I thought was coming from the same, so I followed this way, but actually this was coming from here.

J

This was coming from here, so I have a false positive. That's how I end up falling so one guy Christian close. Yes,.

M

No just maybe kind of true to uplevel this right so before I think maybe doing all these decisions of drafts and all this stuff, don't we want to first have some. You know rough characterization of the resource requirements that we get with the different options right.

G

Yes, really: what's.

M

The you know state that we need intermediate. What's the overhead in the packet and then I guess with the bloom filters. Also the stochastic analysis of the you know: redundant access traffic, I level, I know what is necessary as a first step for this to you know, make decisions.

J

For the next step, right and and and there is a side aspect to that about bloom filters, yes, in the one hand, this fits the bit the bureau architecture. This is something novel we may have to talk to the bureau working group and see if there is a wider interest and how that fits that which I see people, don't they are how that fits. The the the the bureau architecture if you're in general, is interesting. Looking at that, so all those points- yes I mean Taurus computed.

I

Yeah so I'm not sure we will be using the MPLS encapsulation, so maybe the the ripple form of this is actually different from other forms began. As I said, we are using the rank. So this is the constraint cast proposed that pieces closely tied having ripple below so to compare the the various few things with metrics, which I very much like to do the the bit-by-bit Ching and where that doesn't really have to be. There are other good ways to compress it so.

I

You have to number the things that make decisions here, so you're either number hosts in the story mode column or you number forwarding notes in the non story mode. So that's one one different well.

J

Well, you will have to decide what we number I mean. That's about.

I

J

Yes, the amount of information is correct. Yes,.

I

So we need a protocol that does this numbering and I have expressed some some concern that this particle will will need some robustness to actually work in a dynamic.

J

Yes, I mean there are answers to that. Well, we already give away shot addresses in those networks now all networks, so so that we can and over a bit at the same time, for.

I

Short addresses we usually have densities of using the number space which are low enough- that we can actually can absorb some charm because we just need.

G

I

Address located for a while and so on, if we do this for the bit allocation, then of course we are not counting the number of hosts, for instance, we're counting the number of hosts plus such as charge factor yeah. That will need to be yes.

J

Internet we had in mind was leaked link that to six dependency, because you actually block is it. You actually register an address, go to the 6 lv r + y lv r, as your address can can map it with a bit and usually it's co-located with the repo route, and you know it has a lifetime. So it knows it's like a lease now it turns over all these bits for this list time it you don't out and that's how it'll be mint is.

I

The easy part, the hard part is when you actually reallocate this bit of somebody else so.

J

I

You need this chance record. The.

J

Churn helps, you know how things are laughs, but you have a lifetime. Yes,.

I

So you have to look at the details of yes.

J

I mean I, understand, I mean that's Chell.

I

Helps so apart from that, we need a couple of fields in a few messages to actually carry out this and.

J

That I think all these goes into Tallis proposed work, which is how efficient are the bits right. In the one hand, you have the numbers that you showed about the the bloom filter, which says: okay with 100,000 256 bits get this 30 percent or whatever I mean those are numbers of how big the bloom filter has to be. The bitmap has to be bigger than the number of hosts because of the churn. So then again, it's not I mean you have loss. So we have to compare all those losses. Basically, I think.

M

What I heard Carson saying was: maybe what I was thinking that the second step, which is kind of exactly this, you know, what's the control plane necessary to do the management of the address space I mean the question is a little bit sure we do just already start brainstorming. At least you know the the rat Megan is not not the exact protocol messages that we need to add to the.

J

Control plane is easy, fast is about I. Think if I understand well, Caston and I would agree with what you said is the hub thought. Yes, I I have 200 hosts. How many bits do I want, because I have true right: I want to keep some some buffer here, so I don't reuse a bit too fast and I. Don't call that control plane. How do you call it well.

G

J

Really, the decision of City the control plane for me is the protocol, which goes and give me a bitonic all.

K

J

M

Common terminology, okay, and of this management of the address space- is, you know another big block and the question is: should we already, you know in the beginning, try to at least lay out what the options are yeah exactly we.

I

Can never the for the assignment. In the left case, we can average ripple for the bit assignment in the right case, so we have enough places to put this information, but we we have to write up one choice.

G

I

Point for that part, and what we also could do is increase the number of rows in this table, because we can look at other ways of compressing. The bitmap and I think that there are some pretty good ways and these pretty good ways may actually get close to the efficiency of room. Fitness, so I think that that's something that that's worth looking at and are.

J

You talking about basically I have a bit by bit assignment and then I compress it on the wire somatically still the same bitmap, it's just yes, because this is completely different right! That's why, for me, these are also.

I

A bitmap, it's just pretty large in the uncompressed.

J

The point being that you're still there but you compress it yeah and and so it's.

I

Not as big so you can put the design numbers into a room fitter which probably doesn't make sense, because the IP address caches as well. That's the exact number, so there's a little bit of space between them that that could be used up to a probabilistic compression which could be used for really not large networks.

I

So that's the the vertical part and on the horizontal part, you know I'm only interested in the right I know.

J

You only only interest in that half for the non storing unicast world I still fail to understand why you extended this condition for the multicast, because as soon as you build what you build constraint cast and you have the destination IP address and you have all the hops because you learned them from so you're doing non, storing mode unicast. So in that world you have all your dresses, you know existing network, and now you know how to get to any IP address at the leaf.

J

You know all these and you want to know the path to the leaf. So at some point, I was wondering eh. Why don't we keep what you know for the I mean? Why don't you do an extra from far.

I

Beyond you, the answer for that the the answering case is slightly simpler, because the the actual source is always the route. If you have a saw somewhere in your tree, then you have to work with both the the upstream and the downstream case and.

J

Just cool because you may have a false positive because you have yeah it.

I

Becomes a little bit more complicated, I.

J

I

J

If you're doing, if you're not doing blue, that's not a problem but with bloom, it becomes a farm you're not doing bloom. If you send a packet from down as long as you don't resolve all the bit down, you still have to copy up, but you can still forward with the bits with the bloom you'd never know so so. Okay, it's one thing to to mark in our list of protocols or capabilities, the the bit-by-bit allows you to send a multicast packet from inside the network.

J

It's part of the differences, much easier, ok, so I, don't know if much time we have. We have already the middle of the discussion. What did you have? We have?

J

You have lots of time? Ok, so so, actually we we started discussing what I had asked questions anyway. So what is the protocol to allocate a bit and discussed and said I mean what how many more bits do we need? Then we have addresses. So we can. We can keep. You know this result to use a bit allocation packet. Compression I mean it's not just having the the beer, but we also need to transport a bit map in the packets, so we have actually started a 600 draft on that.

J

It's not very live it's just too sure you can do it. If this group now starts working more actively on beer, then certainly I will revive this work and worship.

J

We have to express, we have to chant ripple which evolution we take. We need to be able to express new forms of dowels, so we express bits as opposed to dresses right but and, as Kirsten said, that can even be compressed a lot actually and.

J

Another thing which is, we always said we would compress repo mechanics, we never did so yes, I mean the dowload, yo-yos and all sense I mean and.

J

And then we've got, we've got how we do do we and oh I've got more hosts and really I feel safe to endure with with my bitmap, what happens I mean: do we make groups, do do we I mean with bluem? You you've got actually benefit off. Okay, if I'm beyond the statistics, I hope that rien creases my false positive, which increases number of packets in my network, but I, still work in a bit by bit. Network. If I have more hosts than they have bits, then some hosts won't be served.

J

I will start to give a same packet to both hosts all the time. I don't know.

J

And for the beauty, then I I know a number of of responses to that in customs document. Actually reading it was very light, I mean troth. Could that's more words in explaining always, but I mean we have a number of answers. What we place as link ID could probably be the IP address of the next stop because I'm not even sure. If that's the next stop I guess it is.

J

Then same question about 600 right, same question about 3po some question about bit: overflow management in one: do what do we do when, when our bitmap is too small, do we need to remember, then our network? Do we have groups, Group, 1 group, 2, etcetera,.

M

Sorry totally clueless, but what's what's the mandatory native forwarding plane? Does it kind of within the domain have to be an ipv6 packet at the outer side? Or could that actually be? You know also a beer packet right, because we have all these. You know overhead right now with you know the ipv6 in ipv6, encapsulation and so I was wondering if we can replace the the outer ipv6 header with it with a beer header. Yes,.

J

Actually it's whatever we like, because we have this six lower edge compression mechanism which already compresses the IP 9pa and dramatically to almost zero right. So it's only implicit somehow you derive it from something. So at some point we this draft has it. We are almost carry only the bits, but we say how you could reconstruct the whole packet.

J

If you really want to show that we stick to the ipv6 architecture and to the pure architecture, there must be a way through all those implicit and what to reconstruct the four IP packet, but we never use it. We actually form on the packet in the compressed fashion, meaning that the only thing you will see mostly are the bits.

J

So it's officially we comply to the architecture because you can always reconstruct it. The packet.

J

So since we have a little bit more time again, I would just show you in a few second just to see if we have reactions here and we have new people, how we thought the unique ass would work and then, but basically the multicast is pretty much the same. But it's very simple. So you know repo builds directed acyclic graph to us roots right. So so, in my case, I have four roots here and the backbone, so you have four directed acyclic graphs and so I.

J

Just pick one and I gave in the tries every node has formed an address. You know your neighbor discovery took off six of em d-block and in non storing just pick the case of non storing. It works either way in non. Storing a node device here. Well tell the root, with a unicast and to a message MF and my parent is B. B will tell the root and B and my parent is see.

J

The result of all those advertisements is that the geotag that has formed as a wave outside of the root is told to the root as a sequence of vectors. So all all those those vectors are expressed one by one to the root and the root is capable to reconstruct the audio deck.

J

So so these vectors are expressed like a target and the transit right target to translate with an arrow here.

J

So actually yeah yeah. It's interesting because I pick the know. Storing when I make this example, for they pick the story.

J

Well. This is past.

J

Well forget about this. This is just ways of allocating the bits that we discussed in our earlier meeting so say we have. We have allocated the bits okay, so each node has one particular bit which is mapped to you said the node ID could be just its IP 1 IP address. If you'd use expand nd, you will register an IP address and for that IP address you will get a bit. That's one way of doing it, and so for each address now we've got the mapping yet right, no surprise for the beer people.

J

And so that's where it all starts so in storing mode. You basically tell your parent here is my uterus and all the children do that the parent aggregates all those addresses as there is what goes through me and tells that with some parent with its own address. Added to that.

J

So that's how you paint the geoduck so ripple as this first wave, which creates the geotag and this written wave which paints it with all the prefixes and since it's a distance like the operation, you only need to know who is reachable via whom right and that results in us, as many reverse geoducks, pointing to all those addresses which are just wearing the same tear duct as the other one, but the other way instead, so that that's what I'm expressing here so now, the half here will be telling be a I.

J

Have this bit here, not a will be telling be a I have this bit here we will keep the bitmap for always children and only children. It will aggregate that with an all operation, so it will aggregate its children and himself. So that will end up with this bit. For F this bit for B, so it should before a and this this it's on bit and that's what it sounds to its parent.

J

So you see, you see three bits now: a F and B, that's the result of the or operation, and you keep orange orange orange all the children and maintaining the bitmap for its object. All of the children right and- and you end up at the root here, which has I picked as many bits as ho. So if you get all the bits set somehow.

C

What's come in case George in case you have too many children and you guys.

J

That's the problem of the bitmap, the management right. What do you do when we can? You know tend to overflow this and that's a question to be asked: I mean in beer. You can do groups all right. You can keep the sames, but you have some gates to say the group, but then, when you send something to different bits which are different groups, you need to send different messages. One per group- so that's one way of handling it.

J

So this is basically telling the story. I just said, so you see the debits aggregating. So now you see this path, for instance, so I and G and M K. This is.

J

Keeps that I hope it's clear yeah, and so, when you have a message and I'm just talking about the message down, but any node which any node, which would get a message to a certain bitmap, could forward it from any place. It will still work.

J

Let me show you what happens for a message to from the root, so the root will look at the one or multiple destination. That message needs to read to reach a unicast is just a special case, and so it just computes the destination bitmap as the or of all the destination that you want to reach. So, for instance, you want to know what example I averaged you want to reach to this nation. You just create a bitmap of this to destination and.

J

Now, as the packet is forwarded, the node will look at all its children one by one, and it will do an end operation, the end operation between the bitmap of that child and the bitmap of my destination. If the end is nonzero, so there is at least one bit in common. That means that this path is going to one of the destination.

J

It's pretty easy operation.

J

So you have an optimization if you find that the bitmaps are pretty, is pretty much on your children, you just broadcast it are supposed to select if unique as one marrow I supposed to be here, you don't have to reset the bits as you fall, because it's a dyrdek right, the bits who can be sent and there will be useless from then on it's an option. We have to decide.

J

If you want to rot up, you will have to erase the bit, and this is an example message to a FMJ. So you end up building this bitmap yeah. And if you look at the slides, you will do the forwarding operation.

J

The interesting thing- and we discussed that last time is. We can also make that reliable kind of, because once the packet reaches the destination, we can send back an acknowledgement and the with a bitmap as well, and the bits can be odd as flies up and the result when it reaches the root is the awe of everything so those all the nodes which got it so so this descending with the send bitmap. The active map gives you the missing. Bitmap, you retry that so that's an interesting side effect.

J

So if you only route for a route from the route, then you can avoid resetting the bits. If you want to be able to send a packet from anywhere, which is the new thing that we just discussed with Gaston today, then you need to reset the bits for NTP architecture.

B

Thank you very much. Pascal I wanted to see if there's interest in this work in this, for this work in the working group, I mean there's already been in considerable effort not to influence. You I think it is very interesting, so maybe I can have an arm from the Virgin Group if they think they will should take on this work. They publish beer. Please hum now.

B

Anybody Fairey against this ID, so please speak up now and give your reasons.

I

Just more, I I don't like talking about things. I, don't know anything about, but there's always.

G

This 800-pound.

I

Gorilla in the room that we don't see well, which is called IPR in the ITF, so that keeping things simple might create a draft that that is easier to control in this space covering the whole floor may be more difficult to control. Okay, well,.

B

One of the sinks that he sought the way forward was to start a design team is in select number of people and that they will look at the work analyze. What should be done and come forward with alternatives if needed, tell them when it involves circumstances. It is interesting to do this. Well, then Republic off it and also to see what kind of documents we need. That was the ID for an design team. Then all the companies.

I

That will send patent attorney into this design chain.

I

B

F

M

That always the case, I know you know I would I would I would think that if we do a design team and then I think afterwards, the thing would be. Okay, if we adopted say this option here. Does anybody know what IP are we somebody has on this? I mean it's?

M

If you think something is simple doesn't mean, somebody else hasn't already tried to put a patent out right, so maybe you're bloom filter. Just because all this stuff is so old, they all expired, but.

M

No but I think it's a fair point and I think that may be exactly why you know first doing a design. He may be coming up with some options and then you know, if we're afraid of IPR, let's just simply have, and they.

B

Can announce they're done? Okay,.

M

Very good, we can't make judgments on what we don't know, but if, at that point in time we we get, you know sometime, we learn something we don't like. Then we can take that into account with the next step. So.

B

That's the idea of the design team to remove the gorilla and get just a small monkey yep.

B

Thanks there were already suggestions for participation in the design team. That's Kirsten, I hope he wants to do. I mean really appreciate it, and Pascal I think you certainly want to participate in the design team to others has been volunteered.

B

So that would be nice me nervous. Oh.

F

I

I'm volunteering Olaf, who actually did the workman, is.

B

Doing, okay, so you're volunteering, but there is indirect pointer to Olaf. Okay, very good! Then there is Greg are showing yeah.

B

Ice from Cisco, so a little bit of work on the on the website, you'll be a working group. So I would like to volunteer as well for the good I.

B

Think we haven't team. huh Yes,.

B

Yep can I ask to list to coordinate the team yeah.

A

B

Then you try to come to a conclusion, see how the tiny design team looks like and tell them at the next meeting up. Maybe before already have a document in which you say the different alternatives when it is useful to use and what kind of documents we're going to look at and possibly the IPR involved them, is that correct yeah? Is that clear, yeah? Okay, thank you very much indeed, the most difficult subject: I'm glad that you're mocking progressing. Thank you.

H

1.5 millimeters right, Carsten 1.5 there we go they've, adjusted it yeah, I know, I could notice that I could hear them, despite that. So, um since yesterday these slides have been updated so slightly, we had a very productive discussion yesterday in the coat lounge to understand some of the issues. So this sounds like nothing at all what we talked about in six dish yesterday, that's a good thing. It means we've improved just how many people were there yesterday in six dish?

H

Okay, so not not, not even half so the rest of you. This will be a little bit noon, that's good! So this is just a little bit. What we're going through this is basically a set of requirements coming from six Titian to roll to do some things and to do with some specific problems that well, they turned out to be role, problems that we have in six dish.

H

Okay, so they're not specific to six dish, actually they're very much not specific to success, but when six dish we have a specific way that we wish to solve some of the issues. So that's why and we have a weight we we want some layer, two things that we're gonna do so, um which button there we go. So this is like a complicated network that Pascal drew a while ago. We have a blue Network. We have a green network.

H

The blue network has some kind of a backbone happening and has a number of doe tags um in in some discussion back and forth.

H

Yesterday there was some you know discussion as to whether or not these were different, pan IDs at the 15.4 layer and this matters, because, if they're different, pan IDs, then they probably most certainly have to have different keys, and that means a single note sitting here can't necessarily hear this and this network at the same time, because it doesn't necessarily have all the same keys now, there's two aspects to this: there: the problem of of enrollment.

H

So this is where you first join a network and get the keys and do some things, and in that case it's very important that you find out, which ones are the green networks which ones in the blue networks. Because if the blue network is not your mommy, then you should probably shouldn't join it. Okay, but the point is that if you hear a whole bunch of things one from here- one from here here here here- that's a lot of a lot of things that you may have to go through. So it'd be nice.

H

What we would like to have in six station, we have a way of doing this- is to identify the blue networks as being blue and the green networks as being green without disclosing anything to an attacker about what's going on. So the important thing is that if it turns out that you're supposed to join the green network, it would be nice if you tried one blue network and then put the all the other other announcements that you hear to the bottom of your list and tried the green network. Of course, the green network didn't work.

H

Then it might be that the blue network, you saw wasn't route, wasn't with a malicious attacker and you should try a different different way, so you do have to try them all in the end before you can give up. But as soon as you succeed with one of them, then you're you're, you're golden okay. So to that end we are so it's just that's it. So there's the other aspect. I want to go back to this, which is that having joined a network there's then an issue the first time.

H

There's then the issue of well did I attach to the right not to the right parent, because that's parent selection, but now did I tap attack attach to the right doe, dag, because it could be that this guy in the middle would be much better over here than over here.

H

Okay and so then we have this issue that we need to do essentially doe tag selection before we do as part of parent selection, and that becomes a challenge because there may be different keys and there's other stuff like this, and you can't hear the DAO is, and you don't get all the metrics. So there's some issues, there's some overlap between essentially layer, 2 decisions and layer, 3 decisions and we're trying to get the right information to right places so that we can make this thing first, and you have some comment at this point.

I

So if there were in a secret and shaking that pledge and any node in the network could exchange with each other, this all would become very simple right. A.

H

Secret handshake, so in fact this guy would be able to know this guy's without exchanging any information. What do you mean by secret handshake.

I

Well, the idea about the secret handshake is that none of the two parties that do the handshake disclose anything to a third party, but.

G

I

Do this close to each other that they remember ah so.

H

I see where you're going and- and that may be an interesting addition, but I, don't think that our problems problem, that we need to keep secret the existence of the blue network or the green network, or the fact that that this guy and this guy are part of the same network. So what you're would solve is it would allow an observer to not. It would prevent an observer from noticing that blue over here is the same as blue over here.

H

That is not our objective. It is an interesting objective, but that's not our objective. The objective here is that this guy, who is potentially maybe an attacker, but there potentially is an interesting and a non malicious. No that's say for the moment would like to avoid going through all of these blue networks just to discuss that he belongs in the green Network. It's a head of queue problem.

I

We do that with us inside yesterday, yeah.

H

But we don't and we don't have SS IDs in 15.4, okay, we don't know how to select them. Okay, so in some sense this is an insecure SSID for 15.4. If you want to think of it that way,.

I

Generate a random number and use this as an the.

H

Fact, that's that's! What's on the next, so a couple slides in the future right. So it's just it's just you! Don't we don't so the other thing is that in some sense, pan ID is that thing, but pan ID has some other other. It's too short, but also it has some other other operational issues that issue so as I said so this node it may be better. Particularly. You may find that this part of the network is too congested and wishes to move over here.

H

Okay and I'm, not at all going to talk about how the decision is made, but rather how the information about the status of this network is occurs. Okay, so I know, don't ask me questions about the parent selection, evolving, multiple things, because that's Georgio's problem, yeah right, okay, so I just wanted just clear. There's two parts to this was enrollment and parent selection.

H

If you saw the slide, we had some dispute because I didn't really see how parent selection fit into it and I now understand at least well enough to update my slides, if not well enough, to update the document. Quite yet so clearly, the blue and green networks are different. Blue Network, in this case minus three pan IDs and different pan IDs leads to also two different. If you different ipv6 short addresses.

H

So if you have different pan IDs and you do it because you needed more bandwidth with different schedules and things like that, then I think that one of the effects is that you don't get.

H

What was the rigid use? Transparent? Note, painless, no use another word, seamless, seamless movement between the different doe tags. You would have to renumber, as you moved from one doe tag to the other, and that's okay, that's that's like compromise. You may also build. You can also build this network with all three note three having the same pan, ID and the same network, but then you get into issues with synchronizing the the asns and things like this.

H

If you're doing six dish and cryptographic reason, so you probably wind up- are forced to have different pan IDs. If you have overlapping networks. If, on the other hand, um you you this is not the case you're, the green yard, you just build a really big. The green network- and you just happen to have multiple attachments to the root to the core. That may not be an issue. Okay, your your your doe tag.

H

Your DOTA dag, could extend up to something here with this really being a parent and really all being one one doe tag, and you know all one hand ID just multiple e attached, though the highest levels happen to be Ethernet rather than 54. That's that's a different problem, we're not really specifically dealing with so in six dish. We've changed this diagram, probably 12 times since yesterday, or something like that. So we are gonna, create an information element using the IETF allocation.

H

That's now available, it will have something called a proxy priority which is essentially determined by. If you want to enroll. How likely is this proxy going to be able to help you so, for instance, how many neighbor cash entry does it? Have we heard rales, you know simulation was eight right, we know was not. You was someone else previous to that at eight right. How many do you want to allocate out of your eight neighbor cash entries to an untrusted node that you don't know who it is to do enrollment?

H

The answer may be one possibly to for that. Eight I, probably I, probably be skeptical in one. So if you've already got something going on and it's it's continuing, then the proxy priority probably needs to be a very undesirable value such that you don't get any more profit more nodes. Trying to join they'll try something through some other other proxy other hand. You know you've got a hundred twenty eight entries. You know you could do ten percent to that easily and you'd have a very much better.

H

Our is a bit that essentially allows us to say that this node is a a router from the six-man point of view and would accept a unicast router solicitation and the reason why you'd want to do that is because the alternative is that the new node or the leaf node in the if the leaf node would have to either. According to this, the rules would have to either do a a broadcast, router solicitation or wait for a broadcast router advertisement. Okay, so this says: hey you: can you forum with my my lair to address?

H

You may send me a unicast router solicitation to get a unicast router advertisement which eliminates the need for a leaf node to tie up a broadcast slot in a schedule. So that's sort of a separate thing, that's happening in there.

H

We have this network ID, which used to be 16 bytes, but is now variable as of yesterday, um and it's probably something like a shot. A 15-inch ah shoe 56 hash of a your prefix, um but we've decided there's it's set at the DOE tag. Id and that's all. Anyone else needs to know the DOE debt that the root give me the root, sets it and tells people about it.

H

Finally, we have this doe tag, priority, which is the relative importance of the say, the three green pieces, okay- and that said at the root based upon some knowledge it has about the traffic through its part. How how congested is it at the top? How willing in is it to accept a new traffic? That's not just for enrolling nodes, but also for that note that needs to jump. That's already enroll that needs to jump from one network to the other, because the the et X's are too poor for some reason.

H

So that's what six dish is doing. Ok, so proxy priority is primarily a local thing. The Arbit is primarily a local decision.

H

The DoDEA crier tea and the network ID need to be transmitted. That's why we're here Karsten.

I

Let's say I'm in Attica and I want to take the privacy of potential China, so I just send one of these with.

H

I

Id and high priority yeah come.

H

Talk to me yep and the the the node in tempting a to enroll will attempt to enroll, and we have pieces above this in the enrollment protocol that assume lists and detect that you don't have a legitimate claim.

I

On, if you find out that there is a node somewhere yeah who would like to join? Yes,.

H

I

Network yep, that's information; disclosure. Yes,.

H

We agree: it's information- disclosure. Yes, absolutely! Ok! So we have this problem right at some point. Somebody has to speak first, ok for people to find for people for note for me right. So if you want to create a secret thing- and we could do this in a zero touch way, then I'm. All yours like tell me, tell me how to do it. I don't know how to do it, but but it could be that you know something I, don't pry quietly. Yes,.

N

Don't what Dana I'd like to stress that, so we had this discussion long discussion yesterday about this. Is this working? Yes, yeah and and I'd like to stress that this is a. This is a problem that it that I face.

N

You know very often the world upon these things with my company, you have different that different, essentially doe tags that are deployed in the same space and a new node has to pick which one is joins, and so without something like this, it will try and try and try and try and try until it hits the right network and and that's a waste of energy. It's a waste of time.

N

It's frustration for the tech guy, so so this is a way of announcing with the network what we call Network ID, announcing that you're attempting to join. What you believe is the right network you may join.

N

You know you may join it as Carson was saying you wait, you may join the wrong one for the wrong reasons, but joining will fails. We will try to write one death words. You know with the proxy priority among the different doe tags. It's actually the same prefix, essentially which ones you should prioritize over other ones. You don't want to attach to a network, that's completely congested and then have another network right next to it. That is completely empty and- and you should attach to that empty one, and then to do that.

N

Priority is something that I'm sorry. The proxy priority is something that you know is more more local, but I want to stress that this is coming out of the real use case. It's not something you made up.

J

Let's get you back, I can only agree with Thomas, because our whole networks are hitting the same problems and it's it's more than probably the one I'm most interested in and is the DoD priority, because, basically, what that tells you is when you see multiple geo tags and that's true for the joint process, the enrollment and that's true for the reload, the life of the device, the device will see multiple do-dads and what we see in the field is that there is an instability in the structures that we build.

J

For instance, if one vo dag has a lot more devices and traffic than the next. But when happen is, is there will be a lot of load on the radio which will create loss which will force the ETX, which is one of those metrics? We use a lot to to to go down and as the the ETA goes down, then the rank goes higher and all of a sudden, the dog. Next to you, becomes more desirable because the rank looks better there.

J

So notes move like as a bunch to the next do DAC and guess what you create. One of those associations and that that's something that we actually see in the field, so it's not just about defining and do that priority. It's about defining it well about expressing it's it's something which mostly hits the root right.

J

The load at the root is mostly the load of the network, so we have to carry that of our repo from the root down to the devices, and we have just carrying a metric is not that hard right just at this yell in the DI? Oh, but it's! The real problem is this control loop, which will stabilize the to do tags one next to the other, so that you know that they will end up in a situation where they are evenly load, unloaded or something, and it's even defining evenly load.

J

It is not easy, and so the work we have to do there is not just to pass. Two bytes in the I/o is to understand how we can end up with a balanced and stable condition for the runtime of the network, using having ranked like that, which is a geotag level rank.

H

So so the the last thing is says that there is a we had a rank priority, which is was really that the toe tag rank, and then it was pointed out to me that the other information element in our in our beacon already has that number in it. um So we didn't need to repeat it, but it was a news to me. So I wasn't familiar with out of her, so this just sort of just describes a little bit the things and I used. The word I said preference.

H

So I didn't made a mistake again: I try to consistently all I. Did it all here. That's what I get for doing it. It 7:00 a.m. this morning, I mean to say where you see preference bright priority and trying to consistently say priority where lower numbers are better. Okay and preference is ambiguous for many people, the term. So we what's the role part.

H

Okay, so I haven't updated this document yet, but it has some of this, so we want to have what I believe is a new metric option, new metric container, um which would contain the dough tag, preference I'm, pretty sure it needs to be a metric because it needs to show up in the DI and it needs to be refreshed, go down the thing whether some other objective functions would do anything with it.

H

The answer is no, but as a metric it would trans be transmitted unchanged from from one layer to the next okay, and then I've put this in here and I. Think that's just because it was in the in the previous thing and I'm not sure it has any meaning, but I think I think it gets removed.

H

Now it was there from before and then I think we have a new d io configuration option that does not change much or ever so it would go in the d io configuration option which is sent periodically, but not necessarily with every d io most networks such that this network ID, however big it is, is there and I'm saying probably it should be limited to 16 bytes, that's my suggestion.

H

Pascal said you know, maybe one byte will be do X are together at one point, but what we, what we've, what we've agreed on and the end is that it will be one to sixteen bytes and we already have a length to deal with it so that the the overhead of making it variable size, probably not significant, and that's my suggestions, charge if it's truncated sha-256 to the PIO that should be set in the route, but that it's up to the route to decide. It's not. No one else calculates it.

H

So we need to send it through the network. It's too pretty small things.

H

So we did this goals and sixth issue was to write it down, and a sixth issue will probably adopt this I hope document to figure out what to do and the goal and role is to determine how the newly exposed metrics interact or derived from vio things. Most cases there we've decided they're, not they were derived from things that you'd find the PIO. For instance, at one point, the network ID was going to be the DoDEA ID.

H

Okay and we've agreed yesterday that that's not what we want, because the different doe tags would have different dude do-- tag IDs, and so that would be an incorrect thing, but the advantage was that we don't need to transmit it if it's already there.

H

So people number of children there's a lot of other things and we haven't. We haven't expressed that, although the suggestion yesterday was that the dough Deng pref priority could literally be how many children you have in the hold owed, how many members in the hold oh dang, because as it gets bigger, probably the desirability of that dough dad goes down, which is pretty much exactly what you want.

H

Of course, you might not want to disclose how many children you have- or you might not want to tell anyone that that's actually how you make the metric, but you know that, maybe that's fine I, don't think we need to so we had some discussion about whether or not we need to standardize what the values are, because if you have a dough day route, which is from different manufacturers, then they may do the calculation differently, which means that this, the the values don't have the same relative value. So that's something to think about.

H

We may actually need to specify an algorithm there. Okay, I'm, not I'm, not I, don't know, but what they're? That's it? Of course, what I would like is each adopt my this document here or what successors like use you dot this document here. So any questions.

J

Yes, I just want to react to what you said about this metric. If you look at it, the way you compute this information is pretty is very close to what the normal objective function.

J

Does it's just not doing it for one node is doing it for a graph, but it takes an account, maybe the number of nodes in the network or the capability to get more nodes, because if I have like a route which which can have 1,000 nodes and a route which can have only ten and one as the one with one thousand is 100 less more entry and the one we could do ten as just two entries right, a ratio of what you can do is not the right value.

J

The amount of how many you can still get in is more what what you are interested in, so the actual number of nodes in the network. What I'm trying to get is not interesting. The number of nodes that you can still take in is probably more interesting. The percentage message.

H

Is were successful.

J

Percentage is the wrong thing. Worst thing you can do right, because if my to do dogs are very different in capabilities, a percentage of 10% left on the left might be one guy and the percentage of 10% on the right might be 100. Guy I want to turn the right. So so it's reading the number of what's left really but oh, and that can be discussed by the world.

J

What I'm saying is you're, taking a number of metrics like number of nodes left or something my amount of bandwidth left I mean the observation of the bandwidth being used around the roots. Usually is what you care for, and, and maybe some other things which come from from the nodes right I mean if, if all the nodes in the network have their neighbor cache saturated, there is nothing that we can do, even if it could take more notes.

J

So it looks very much like putting a number of metrics together into a number again like an objective function. Does so yes, I agree. We have to specify that at least one objective function and- and it would not be an objective function to compute Iraq. It would be an objective function, compute and.

H

It won't be communicated, it will be configured on the Dodig route right.

J

H

And as a matrix.

J

That it needs of Lee I'll, just look at local to the root, but it might be that's some objective functions need information from the Leafs or whatever else I, don't know, I'm saying a nut casting in stone. Just like a row.

H

We don't know the the loading right now in storing mode. We do not know the loading of each of the children non storing mode. We do I know that some people have would like to propagate that information up, and so maybe that's a there's, a connection right there immediately to doing things. Any other questions concerns any other objections to this work. You feel this is like the just totally the wrong thing to do.

B

This is beautiful, most oh cool, I'm, sure I'm bit slow, but I didn't quite understand what you're proposing guys the parent selection included. Is the objective function included, or is it just in the I/o description? I'm, not quite sure you so.

H

It's just the DAO description of a container okay one's a metric one's a configuration option. um That's that that's the the concrete bits on the wire thing. The document has to do its the docking. Okay, then the document, probably is we just said- has to say something about the the canonical way of calculating this okay, or at least for the that that there is a and- and maybe even has an AI n a registry for the names of these things, so that we can actually talk about them. Someone will write a new document.

H

It says a smarter way of doing it at some point. If that makes sense, and clearly, if someone writes a I, don't know a yang module for dough tag roots, it would have to include the name of this. This dough Dec priority objective function in it so that it could be configured, but it doesn't have to be transmitted in any way. The way that our current objective function is is is described and.

B

The parent selection is in different documents and also the order of joining a network and parent selections in different.

H

Different document, so we're not dealing it all with parent selection at all. We don't care about that book, but we do we it's. We don't care about the mechanism, just that, so you assume that there's some parent selection that can get better so.

B

We're talking about another reduce document, there's a clear objective and it's very.

H

B

H

Not a lot bigger than the then the you know, then this kind of stuff to to diagrams like this and I Anna allocation for the for the the types- and you know, probably you know five pages of oh. Why are we doing this and security considerations and whatever.

B

Good, the anybody read this draft.

B

Are people willing to review the draft later on? Do we think we already should adopt it now or wait? The next version.

H

Mailing list, good yeah I think we should wait to the next version and then I hope will be adopted. Quick, but I I need I, haven't even done this stuff. That was, we changed yesterday.

H

She has paper copies of all that RFC's here.

A

Okay, open mic: some additional suggestions, comments.

A

Please sign there: blue sheets, daddy.

E

Hello I'm Aris again, so this is my first time here and I find all the problems that are being described very interesting. One thing I have to notice, though, is a regard replicability. So a lot of the problems you describe like the problems with joining or the balancing of the trees and the churn I have to say: I, don't know if maybe there's some repository I don't know, but there doesn't seem an easy way it to be an easy way for us to replicate these problems in and be able to work on them.

E

So it's very nice describing them that you have real world networks which display these problems, but at the end of the day, since I, don't I'm not able to replicate the same network in a simulation, I can't actually attack the problem very well. So I was wondering whether there would be a way to share some of this information.

E

I understand there are some confidentiality issues, maybe if you're working in companies, but even like that I don't know, maybe there would be a way to share just enough information so that we can replicate some of that in our simulation and try to to work with it. Rahul.

G

Shadow I'm glad that you you're talking about it so replication is a big issue. Yes, so so we we have been working towards the simulation framework ports, achieving something like that. So it's that you just for example, I face some problem. I should be able to send you just a configuration file and you should be able to produce that problem yeah exactly. That is a work in progress. We call it white framework, so we're working on it and it exactly tries to achieve what you're trying to describe you. Okay, that's.

E

Good, that would be very helpful, actually yeah. Thank you.

H

Michael was the name of your framework.

H

Whitefield framework, that's cool. The other thing that I was going to say is that I guess this F Interop stuff is, is coming along to be able to do things remotely, so it would also be I. Think really cool I'm totally with you by the way and we've been through this over the whatever eight years of the distance of the role working group night is longer than that. Maybe ten years um kind of repeatedly where we've had people say a researchers say this doesn't work.

H

This doesn't work and here's all the reasons and, and then the working group has said this is really interesting. How can we, you know, replicate it so that we know whether we've fixed it or not, and then said? Oh I can't tell you any of that so well, maybe you shouldn't have bothered you kind of wasted your time right, because if we can't fix the problem- and you can't you can't talk about it- then I I think it almost. You shouldn't even bother, especially.

E

In some of these issues, which are you know, we can just- you can't actually just use a toy example to replicate.

H

Them and film it's a rational data right and the problem with kouya is wonderful, but it's mean last time. I looked it boots, kontiki right yeah, so you can't replicate the problems that rule was talking about. We're Contiki versus riot can interoperate easily to.

E

The extent that you I think which actually allows you to load any firmware. So if you're.

H

E

Compile to firmware, which is simulated, then.

H

There are some restrictions, yeah.

E

Yeah, that's their own pile.

H

E

H

No, no it in many cases it has you supposed to compile it kuia actually loads, an x86 file on your machine. It doesn't actually do a machine emulation, I'm.

E

Sorry, I'm not or whatever yeah.

H

So yeah so cool yeah. It actually does a kanay t'v build so that the code runs low, runs natively, but but it surrounds it. Looking like it's the target architecture. Now there are some. There are some improvements to this, but that's typically how you do and that's the only way to get simulation to run quickly. Okay, so so there are some restrictions of what you can do and and what I hope is that something like F interrupt will will give us a better a platform on which to run all sorts of native firmwares.

H

E

N

So yeah I'm I mean that's. That's that's the game right! There's a lot, a lot to verify, there's a lot that you know the changes all the time. So there's two action. So there's two project I'd like to mention so whatever drop thanks, Michael for mentioning it. That is aimed at ensuring interoperability between between implementations, which is which is one part of the story.

N

I think you're talking about performance and replicating things like this, so yeah here is working on a simulator, the six dish simulator now the Python one right, the Python one yeah yeah that was released yesterday. One zero, zero one gesture day and Maddie is working on a web interface for that, and so the idea is that, of course, its success at the root at the bottom, but then there's ripple and scope and co-op all that and secure join.

N

That will allow you to enter a scenario and and and get graphs or that will give you a prediction of the performance of your network. It's not the answer to your question. There's no answer your.

A

A

And their concern.

N

Thomas, can you introduce tomorrow session, yes,.

A

We tomorrow we are going to have a session as well in the morning, and this going to be the agenda for tomorrow, so from 9:30 to 11:30, we have two hours and just want to present like LD be charlie and the pascal's want to present his proposal not leaves and then Raoul with the no partner modifications and then as when now p.m. Pascal with a doubt projection troublesome, and what about this modification is going to be like a new version in and that's it for tomorrow. So we hope you can attend as well.

A

We can get your comments during the meeting or, if you cannot attend, please join remotely, so you are up-to-date with what's happening here. Okay, thank you very much. I.

B

Personally, should welcome some feedback on the lengths of this meeting, because usually we have much shorter meetings with very well ten minutes per person and we cut it. So we have higher interests. People keep awake much better, but now we have an longer discussions which are which good results, I must say, but if people will be kind enough to at least tell me personally, if they prefer the longer ones or the shorter ones, I will be very happy to receive those comments.

A

Thank you very much have a nice day.