DASH Community Weekly, 17 Aug 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Workgroup Community Meeting Aug 17 2022

Description

Keysight/Mircea Q&A
Discuss keepalives in CPS test
Volodymyr Mytnyk - SAI PTF update
PR Reviews
193 Dash Outbound pipeline packet test - needs Marian and Reshma review
187 Default drop support
194 Fix inbound VNET
190 spell check

A

Meeting um for august 17th and I'll go ahead and present my screen.

A

And let me know if you can see yes: okay, yeah, okay, so last week when we went over quite a bit and we had an awesome, huge presentation by keysight and mercha- and we talked about these two pull requests here with the config generator and the hero test update and we I sent notes. But we talked about all these different things. While we were talking about the presentation- and specifically, we talked about the need to talk about whether we need that keep alive in the connections per second test.

A

So I was hoping we could talk about that a little bit more today or if anyone had thought about that. In the last seven days um I was hoping we could have a short q, a on that, maybe starting out with that. If you wanted to do that mercha and then I was also hoping we could talk about- maybe moving towards stage three of the test maturity stages.

A

Does that work? For you guys.

B

Yeah for the keep alive, so there are two things that we I mean we can do it's either. The flow timer could be modified from one second, let's say to two seconds or three seconds, and then the keeper live is not needed, but then it changes. The fact that hero test is keep the flow timer at one. Second, that will be a possibility not to send the keep alive, um the other one.

B

Since the data was already sent, um we can basically now send the keep alive and then the flow will be terminated by the dpu.

B

um I don't think it will affect too much the stats on my side, but probably the dpu will consider these flows as terminated if it holds such stats. So yeah.

A

Those are the two options: yeah did anyone need.

B

A

Of what was going on there with that conversation or they didn't attend last week, any questions on that.

B

Yeah, sorry, for that I can give a quick summary. So, basically, the moment uh device has a capability of, um I know- let's say 3 million cps, but the requirement is: keep six million active flows.

B

You can bring only three million up every second, so for a second, you bring three million ups on the second. Second, you bring another three million for the total of six million. Now the first three million flows they need to leave for two seconds. Otherwise, if you terminate them on the second second, you will not have six million flows uh on the dp. You will have only three because you expire the other, so that means that the flow needs to leave.

B

Let's say two seconds in this case, and if you need to leave it for two seconds, you need to send the keep alive or two. Otherwise the flows will be expired with a timer of one. Second, so that's add an extra packet or two, because if you actually need to keep it for two seconds, you want to send the packet at 0.999 seconds kind of uh and not wait for the second to pass, because otherwise you know the flow is gone, expired by the dp.

B

So that is a context and that adds an extra packet um and where the discussion started is because the hero test specifies six packets and talks about the tcp. Packets um does not mention anything about data and neither about the keeper live when it does the math and yeah I believe there are. There will be data packets, which will be two packets, so six plus two plus a keep alive.

B

So we are already at you know eight to ten packets for each flow. At this time,.

B

Like I said, the solution is one yeah. We can send the keep alive. That's one solution, the other will be increase. The flow timer from the test from one second to two seconds or three seconds um and the other one will be basically don't send the people, even just let the flow be terminated by the dpu and.

C

See what happens.

A

Does anyone have strong opinions on this? I know um we talked about it a lot last week.

D

And russia, um just one additional suggestion in term of just leaving the timer at two seconds. Can you just add some variant to it so that we can have the flow aging out at the different time so that, basically you you will handle the case where uh new flow is being learned? Why the existing flow uh also being uh time out so basically have the capability to vary in the flow time between one and two seconds.

B

Oh yeah, the flow timer can be configured. uh The discussion was only about the fact that the test, as it is described in uh the hero test document that is present in git, says one second, so yeah.

D

B

D

C

B

Yeah, it's it's configurable uh people can put it to two seconds three seconds or one. Second, I was just bringing it up that either the number of packets increases or either we need to increase the flow timer. Both are configurable. So.

D

Yeah yeah yeah. I don't mean to ask um any change to the approach that you mentioned. I meant to ask, even though we configured like one second or two seconds, um should we allow the configuration like uh within like between one and two right, so that for that three million flow you can vary in the timeout.

D

So it's simulate a more realistic scenario. That's what I'm trying to say, because in the real network, you're gonna have three intermittent flow, all coming once and all each at the same time right. So that's what.

B

The test requires that to have all these flows aging out and new flows being installed every second- uh and this is mostly like a stressful scenario- it's not let's say the real use case. I think in the real use case, this flow timer. It's at I don't know 60 seconds 90 seconds. I think it's set much higher, but in order to stress the device, the most stressful will be basically one second and basically have new flows added and removed every second. So this is just a stress scenario: it's not the real use case.

D

Yeah, but the same test you written, can also be simulated to a typical scenario as well right if you allow the variant of the timer rather than the fixed timing for the flow that that's all I'm suggesting.

C

Yeah, of course,.

B

D

Got it? Thank you.

E

Hey, uh actually, I have a question here, so the flows that you're talking about here um is this the background pro you're referring to or the cps flows.

B

There's a cps flows.

E

So for cps flows uh at least my understanding is like a six packet flow right like so. It just opens the connection and then closes like. So why do it has to stay for like more than a second like, uh like it should get close within a few microseconds.

B

Yes, but in if you do that, you will not have in the flow table the six million flows that are required for seven million flows whatever with. If we consider udp traffic as well.

E

You know like that is a background traffic, so.

B

You have one million background tcp, you have 1 million background udp and you have 5 million, which gets created and destroyed every second.

B

The problem is: if the dpu is not able to do 5 million- and it does let's say 2.5 million, uh then you need to have the flows live for two seconds in order to have a flow table installed in the dpu of five plus one plus one million. Otherwise, your flow table will be only it. It will be very small. The flow table will not get full if you don't keep those flows alive.

E

Now, that's the part I'm not getting like so why? Why is that a focus right like so? We have a background traffic with uh like as per the euro test, 2 million tcp and 2 million udp background traffic and then whatever cps, that the dpu can do like. If it is, let's say, 3 million cps, then off of it will be tcp, 3 million. Sorry 1.5 is tcp 1.5 s udp, so the flow table usage will be like 4 million for background and 1.5 for udp.

E

Since we have one second timeout right, then the tcp will just at any given point of time. Maybe there will be like 500 to 100 active flows within the dpu, so but that's the expectation right like so uh yeah like I I'm not getting why the uh usage of the flow table within the gpu is a focus right. So this this is just working as per the expectation, so, uh like the connection, will just come and go uh because it's going to stay alive only for within those few microseconds.

B

Yeah, so let me share my screen and I'll show again the paragraph that is in question flow table size you need to have 14 million float, work, 14 or 7 million. When I talk it's, if it be directional or not, some, it's calculated uh a 7 if you have both directions or 14.. So um this is what I'm talking about flow table size I need to see in your flow table or the device flow table. 14 million flows.

B

Okay, you have two from tcp two from udp background traffic and these are from the cps ones. If I am to keep this flow table at this size, if the dpu cannot do 5 million cps, I need to leave the flows longer. Otherwise they go out of the table and the table never has 14 million size. This is the line that I'm referring yeah, but.

F

Maybe this line can be changed right. I mean it's like this is what I'm saying.

B

Given the age timer can be changed, I mean if we make changes. That's fine but yeah.

C

B

Be agreed with everyone so yeah, you know we can change here and put two seconds and then it's fine, maybe, but then, if somebody cannot do 2.5 million cps and they need three seconds and again the test changes. So I'm okay with changes. This is why I'm saying there are options, but there will be changes and somebody needs to agree mainly basically microsoft. I would say so.

C

B

F

Change the test, I think the problem with this. The way this is written is that there's no explicit description of the flow lifetime, like you're inferring the flow lifetime from this line in this in this uh bullet, and I think it would be better if this would like explicitly you know said the flow lifetime should be x um and then you know the flow table size is what it is based on, whatever the flow lifetime, whatever the flow lifetime is it's I just. I think it's like the specification of this hero test is.

B

Let me give another line here where it says: each flow must receive one packet every second in each direction, in order to keep this and fill up the links. So if I need to send basically this.

F

B

F

C

F

B

I mean you need to send it also for the other one. If you want to keep that flow table, this one keeps saying.

F

But but this bullet is referring to the background traffic, because it's saying one packet every second and then it says in conjunction with the dynamic traffic.

F

Right so like, I think that this bullet is really just referring to how to keep the background traffic uh alive.

A

I'm taking a snip of that just to make sure we can get clarification. John.

B

Yeah I anyway, I.

A

B

Don't find a way without sending the packet to have this flow table and actually ask the dpo hey: what's your flow table size and I get really small flow table sizes? If I um don't yeah, I I did initially the test which I was terminating and sending the fin and everything super fast. um But then this flow table size. When I asked the dp what your flow table size was really small, like 10 000 200 k, not even reaching a million.

B

I'm disregarding this background traffic, I'm talking about only this, because this is in questions, the background backgrounds that stays there.

E

Yeah, so if you see here right, it clearly says tcp connection is established and terminated without any data packages, so cps tcp connection has to be six packets right and, like so like uh john said, like I think like uh clarifying the lifetime of the cps connection would uh help, uh but, but as far as I understood like uh it, just does the six packet transaction and it just disappears, but, like you have off of the cps connection at udp right, so you will still get uh like in this 5 million cps you.

E

You will see a 2.5 million connections from udp that stays in the flow table. So it's for the tcp, where it just comes and disappears uh within this short period right like so yeah. I I agree that, like the clarity, it would be good to call out the lifetime of the tcp connection.

F

Yeah also one other comment on this bullet, like it's not clear from this bullet, whether it's saying this flow table size is like the size of the provisioning of the flow table, because this is like the worst case, like number of flows like based on a one second timer with that connection per second, it doesn't necessarily say that the flow table needs to be filled to that size.

F

um So, like I, I just think that that bullet is problematic, because it's not clear- and I think that you're inferring from that- um maybe things that didn't necessarily need to be inferred.

B

That is correct that I'm inferring from this text that I need to fill the flow table size to that and if it clarifies that the flow table size when I pulls the dpo stats, doesn't need to be 14 million. Yeah.

D

No but john, we do need to test that scenario right at how the system behave when we fill up the flow.

B

D

B

The way the flow table size per the scale document, it's actually around 60 something million flows.

F

I agree I agree, but I think that that should be like that should be tested by um specifying what the flow lifetime is in conjunction with the aging time you know, and and so like you specify those and the flow table fills fills up like I think I guess my biggest issue is like.

F

I don't think that you ever need keep a lives um for for the cps load um and I think, like you're inferring, the keeper lives because of this bullet, and I think it's just wasn't clear enough to say, like you, never need keep a lives. If you really want to test like a full um table that you would control that with the aging time and and the flow lifetime.

G

Right? Why uh yeah? I was going to ask the same thing: why we need to keep alive, because, if yeah, that the device can edge it out soon as possible and the capacity will be tested and the flow table size will be tested accordingly,.

A

I think it was because the flow table was quite small. Wasn't it mercha.

B

So if I close the flows as soon as possible, then when you ask the dpu for stats what your flow table size, the return value is very small compared to this. So the way I'm inferring and I rendering this it's flow table size, equal, 14 million, okay- and this is just the map how they reach to 14 million. So when I do a dpu get flow table size, I need to see a number close to 14 million during the test. This is what I'm inferring this map.

B

I'm you know this is just the map, how they reach the 14 million value. But when I get the flow table size stat, I need to see 14 million reporting.

F

Yeah- and so I I think, you're assuming that that's what this means- okay, but I.

C

F

I interpret this differently. What I interpret this is to say that your data plan should be provisioned with a flow table of 14 million flows, and that you know the dynamics of the test will determine like how full it is out of those 14 million like like when. If someone said, what's your flow table size, hey, I've got a dpu. What's your flow table size, oh, our flow table size is 64 million. It's like a static number of your capacity.

F

I don't think that this is a statement as to like when you run the test. Given all of these parameters, you have to see 14 million. I think that you can adjust the parameters to show that it gets to the 14 million.

B

Okay, in that case, this one, I'm now passing it over to microsoft, to adjust the parameters or clarify here.

F

B

You know you understood something. I understood something else. Somebody.

F

B

Understood something so.

A

B

You know what it is.

A

Yeah exactly yeah exactly and that's good that we can run this down so we'll try and run this down for you. But you know: beware, I might send you guys an email just to clarify like okay, john okay, restaurant mercha, like our sores ned, I'm just going to call you spread ned. Are we all on the same page as to what we mean here when we're taking it to the scn team? You know as to exactly what we're testing so and gerald. Of course, when he comes back so.

C

A

You see an email for me, don't send it to junk.

B

Okay and in the lab just to be clear, I try all variations here. I try basically don't care about this value, just let it go as fast as possible and see what it is. But then I don't see this or increase the flow timer, and then you see this when you see what happens so yeah it's for me, it's very easy. Just we need to decide which version we want to do.

A

Yep yep: oh that's great help you guys! Thank you. It's good to get it super super crisp. So thanks.

A

And mercy we we have a another meeting on monday too, where we can talk about it.

A

I'll probably send out a note saying what I think I'm hearing the problem statement is to you guys and then you guys reply back and refine it, and then we can chase it down. Okay,.

A

Sounds good, okay, okay,.

G

um Christina, um I think we had a small update um about um the sciptf framework that we had adapted to dash and that's now being used in the ci that chris had so maybe voldemort met nick. Would you like to talk about it a little bit question? Is it okay.

H

Yeah, if, if you don't mind it like just a few minutes, probably to to describe what what the change was.

G

A

Were done myself.

H

Yeah, okay, so just me share my screen.

H

uh So I want to show the one of the pull requests that has been closed very recently.

H

Second yeah, so this one, so what we did here is the we changed ci and we talked with the crease uh about this above the change. So we change to ci to use the development actually branch which includes the changes for the ptf would be down that we done for the dash and yeah so right now, anytime, you, you run the view. You create a pull request.

H

The same will trigger on base on that branch, and it means that you can now start using the site ptf to write your test case and use all changes to be done for the dash and uh yeah. So this is the the one of the also development branches that we are going to create the pull request for this. So we we put the example of the overlay, let's say scenario for the v-net, so it's just an example, so it still needs to be updated based on the latest.

H

The site changes, but anyway you can, you can see and how the ptf structure may look like and how to write your own these cases. So we we have the the test case itself.

H

We have some base basic stuff that needed to to run it, and you can, you know just to see how it looks like uh what classes you have to use you know, and what?

H

What stuff you need to do just to set up and tear down your your setup and to be able to run the test case you can. You can use the document that is referenced in the dash repo, so there is a ptf user guide, so you can use it as an instruction to run your ptf test cases. For example, in our case it can be the cybernet and actually that's it. So just you know the brief introduction into the ptf test cases how it's going to look like.

H

I know that there is some test case in the vignette folder right now, uh so probably we can uh during the time we can merge them into one test case on house and them separately.

H

I

Vladimir thanks as you and I were talking friday about this, um it would be nice to update the spy ptf user guide to be a little more practical towards like a dash use case, because it sort of has that the user guide has elements to maybe a vs type, sonic vs type test or a basic switch type test.

I

um But it doesn't really talk about like how you do an actual dash dp test. So it'd be nice as time goes on to document a little more practical setup and usage.

G

And in the user guide today I think we have the rift test, which is the router interface, uh very, very basic uh test. To start with, uh just to you know again, the user guide is basically saying how to use the test framework and how to integrate the test cases um into the framework and all that how to run it but yeah. I guess you know once we have the v-net case solidified the test case, we could uh possibly use that there as well or create another.

G

uh You know, uh example number two uh to show the dash.

I

Yeah in particular I'm talking about like how you'd like a little diagram that shows actual connections. You know right now. The user guide leaves a lot to the imagination. You know it shows an example of a command line, but it has like the ethernets in there which you wouldn't use for a real dpu. You can use ethernet and nick ports just something a little more practical, so someone could actually follow the instructions and just get get an example. Working right now, they're still missing missing pieces in the instructions.

I

That's what I mean and vladimir- and I talked about this, um so I think he knows what I'm talking about. You know it has helped. It has examples like for a sonic switch or yes, but not a dpu.

I

So if we could just get something a little more concrete to help someone actually do it step by step, it would be, I think, it'll be valuable.

I

You know over time, um but I agree that this uh pr was really important, um just just to sort of provide a little more perspective.

I

I was also mukesh from amd, also found out that the um the ptf ptf uh repo, that was in use in side thrift was not supporting vxlan properly and he came up with a patch for that. But then talking with vladimir, he had a better long-term approach and that's to update ptf itself, and so this pull request. He just did um and we committed it uses the development branch of psy, the dash branch, which has a development branch of ptf to fix these vxlan problems. So.

G

Right, I think uh voldemort nick had uh come across this vxlan problem during um the v-net uh test case yeah that was being written, so he had made.

I

That acceptable, yeah, yeah and mukesh founded the same thing doing his v-net tests. So now we have a consolidated solution.

G

Right that is uh yeah reason why we wanted to talk about it today is because we are seeing uh you know more usage of this framework, so thanks very much to mukesh and makesh. Please reach out to voldemort mitnick myself and chris. If you have any issues- and uh you know we can probably everyone work together as christina was saying, you know, uh for the third stage of the measures test maturity, um we could start to use this yeah.

C

uh I had one question: no, this is okay, um so I I thought we fixed it by actually uh moving to the latest version of psy ptfe like moving to zero nine three.

C

So why do we need to now switch to a development branch? I.

I

G

C

Did solve the issue.

G

uh So yeah volumetric, you want to talk about it.

H

Yeah, so I can explain briefly so uh the one the issue that you faced. This is only one issue, but once you're trying to send the packet and compare these long packets, then you can, you may hit the problems so to fix it. We need to patch the ptf again. So this is why we provided the development version of the ptf and probably we're going to upstream it yes. So there is another problem.

G

Basically, has uh the you know, framework that is adapted to dash that's the main reason, and it's not yet there inside, and we wanted all of us to use it before we upstream it to site, but we could start the process in parallel.

I

Yeah mukesh, if you look at your pull request 193, I added some detailed comments. Last night on this.

C

I

C

I

It turned out, I was working. I was working both with you last week and then vladimir, we had conversations and we put it all together. I think this is the final, better interim solution: yeah yeah, but thanks for you know both you all your work on this and we're all finding these uh little land mines and uh fixing them.

I

So um do we want to spend a couple minutes just looking at outstanding, prs and and try to get some motion in a few areas. We have a few fairly simple pr's in the pipeline that probably need waiting for reviews.

A

I

It's going to be.

A

I

Know we have so.

A

I

Yeah we ought to just either say: okay, I approve it or I'll get to it shortly, because some of these have been waiting for a good week for reviews and um pull request. 193 is one of them.

A

uh Do you want a presenter? Do you want me to do it.

I

uh I could pretty much scream.

I

um Can you see so, let's just let's just kind of go through some of the ones that are top of mind? For me, one of them is: is this one that mukesh started? This began the whole vxlan journey uh in this discussion we just had but he's waiting for some reviews. I know reshma, you had asked to review be on the reviewer uh last week. I think last thursday and and mariam's in the queue. So, if you guys could just look at this code, you know and uh sorry here's all the changes.

I

If you want to just take a look through all this now lukas, you need to kind of back out some of these little changes to go back to what um vladimir did just back out a couple things and retry.

I

If it doesn't work, we need to know right away and fix it. You know uh we'll do that and do whatever it takes to get it going, because we don't want a broken pipeline. But there's just you know, there's some changes here that we need reviewing on or just uh you know, approve it and say the mood cash you're.

I

Okay and let's see that's that one that we need to do and this one here also another one from mukesh, that means that's fixing the drop action, so you know either review it say it's okay or you know recuse yourself from the review list, and we can. We can let this go too and um and then I don't know, martin there's one here, but it looks like it looks like we have a problem in the tooling.

I

There was a um sorry there was an action.

C

I

In the site generation there's some kind of a problem here.

C

Yeah yeah, I'm aware of that.

I

Okay, I don't know if this this pr is, you know, pressing or not or just stay in the background, it's up to you.

C

Yeah, it actually is it's. I guess it's, the last change for v-net to unit.

I

Yeah, it would be nice to um to kind of clean up these v-net pr's as soon as possible and get them done, because there can be a lot of energy putting into writing test cases for these now. um Keysight and other people obviously are working on this. So let's try to um close these things up and we can move on being great, that's uh kind of all. I wanted to talk about.

A

Did anyone else have a pr they wanted to.

B

Bring up while the spell check one, that's a controversial one, so yeah. Basically, I had a bunch of spells that are in my prs and then I went and used a tool and fixed everyone's checks in the whole repo, and then I added the git action for it. But the thing is um yeah. If we decide to take this git action, if you have a spelling error, it will get flagged and also, if you use a word that is not in the dictionary but necessarily a spelling error.

B

Like I don't know your name or you know some term that is using dash and nowhere else it needs to be added to the dictionary, and there is a dictionary. The word list file that is there yeah currently has about 600 words added by me.

B

um Is it uh on the left side.

I

Yeah, so basically it's gonna run in the pipeline. A spell check- and you know, given that these are technical documents, there's tons of words that a standard belt checker doesn't know. So you have to build this dictionary um right, and so the implication of this is: if we enable this, then anyone who writes a readme is going to be pestered by spell check failure. Until you add the word to the dictionary or you fix the spell checker yeah, the good news. Is we fixed about 400 or 600 errors?

I

There's a lot of you know misspellings of mine, because I wrote so many documents, but uh so this is.

A

This is a good one. Does anyone have a strong opinion on this? Yes or no, um you want to speak out or raise your hands or you know. I'm.

H

Yeah, just one comment, so I think the spell checking is really great too for the ci, but uh I understand that it will fail. You know if everyone will try to push with some. You know word that is not in the dictionary, but can we make this uh like trigger this job? Do not trigger error but, for example, some warning so yeah.

B

H

B

First thought about to make it as a warning, but git does not allows you. Error fail in progress and pass has no concept of warning and people are asking for this feature for years. um There is some hacky ways to do where you can do it. It will show us a fail, it will still have the red x, but it will not block the pr. So basically, although it's a failure, does not block the pr and you can still merge the pr in uh that's something that can be done, but changing it to a warning.

B

Yeah. A long-standing feature request for git.

A

Interesting prince, do you have an opinion on that.

C

Why the heck measure, if we agree on having the spell check, let it trigger an error? What's the point of having a warning and committing a pull request for errors, even.

A

C

Spell checking.

I

Well, we could try it for a while until there's a community outcry, people come with their shovels and pitchforks.

I

That's a reference to uh yeah hunting, witches right.

A

Right, I would just hate to do it and have people be against it like if, if you're against it, please say something.

B

Yeah, this is a controversial one, so.

C

Yeah, no, I I agree that we should keep the error part because of the fact that you know, if you keep putting the warning in there, we continue to. You know what they call tech. You know, accumulate technical debt right so and at some point we'll have to pay that debt. So would we rather pay it now or we pay it later right and once we have it a lot more, then I think it just basically will be sometimes uninsurmountable part of the thing that we have to really take care of right.

A

Yeah, well, you know the people who contribute a lot of pull requests to the repo we've got bud. We've got john we've got maury and me I do a few restaurants. You guys um bud. Do you have a opinion.

C

I'm fine with it being an error. That's the way it worked at my last uh employer. So, okay, all.

A

C

Sounds like the larger.

A

Sounds like the larger contributors to the repo are okay with it.

I

I'm okay and we can always if we act regret it. We can just change it, but I think it's a good thing to do.

A

I

Keep the standard high.

A

I

This repo has a lot more readmes than a lot of other community projects. There's a lot more explanation of things which is good and it was a huge investment. Just keep the quality high there's a lot of repos. They have no readmes.

I

You just have to read the code and you know we're putting an investment in documentation, so it means we'll make it shine. That's my feeling.

A

Okay, uh anyone else have a pr. They want to check out.

A

I can we can give 15 minutes back also yeah, which I'm sure would be welcome.

I

I'm like we're good.

A

Great well I'll stop the recording and we will definitely try and run down this issue we've been talking about today, um it's important so and john thanks for bringing it up. You know um you'll bring interesting insight, so thank you and you guys have a good day.

I

Thanks a lot christina: okay.

C

I

C