Kong Kuma Community, 8 Mar 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kuma Community Call - March 8, 2023

Description

Agenda:
- Kuma patch releases https://github.com/kumahq/kuma/releases
- KDS redesign
- Q&A
- MSM integration: performance testing
https://docs.google.com/presentation/d/1yG6xgbwbxFGE4YgU8NBGu3RrfDcZyXu4lCccOImiQwY/edit#slide=id.gc6fa3c898_0_0

A

Yeah, hello, everyone welcome to the kuma Community call so yeah. Please add your name to the attend the list and feel free to put any agenda topic you'd like to discuss today.

A

um We don't have much from our site. I to be honest, I would even prioritize some items like. Oh, we have the new here from our NSM, so maybe we can start with that and then, if we have time we can discuss on KDs redesign so yeah the new piece.

B

Oh yeah, hello, everyone uh could I just share my screen. Please yeah.

A

B

Okay, can you say it's right.

A

B

Okay, so um hello, everyone, I'm Daniel from xoret and working on a multi service, mesh project and uh well. This is a project about integrating, in the same with other service measures, uh and one of these meshes is Kuma. uh We have talked about our integration earlier.

B

So if you don't know what this integration is, you can check out uh the database, the record of playlist comedical, which was September 14.. So here's a link- maybe you can copy it retyped it. Oh, you can check out uh our integration example in the comatose Repository, so I I'll give you a short reminder of what this is so and Sam is a service mesh that is based. uh That is level three connectivity, so IP and stuff, and it is compatible with other service meshes and in comma example.

B

We are using virtual Earth L3 Network to provide connectivity with other clusters and virtual machines without using an increase or anything like it, and he has a small diagram. So we, for example, you can have two clusters, both of which have NSM components and on one cluster. We deployment control plane and we can another cluster.

B

We can deploy some service which will have access to this control plane via virtual local network, and so one of our goals with the MSM project was to improve performance because the infinity Ingress components add some latency and maybe some other limitations.

B

uh So in theory this should improve performance, but uh um we didn't have any numbers to compare So. Currently, we are working on actually running the tests.

B

And here's our setup currently we are using uh so so far. We have only one test on using virtual kind clusters and the you have two setups one with Kuma Motors on this is a very basic setup. It's just from documentation. With one exception.

B

We are using Kuma control plane in Universal mode, because we don't have a spare that cluster or virtual machine and in the our integration setup we we just using basic, basically the same setup, as is what is merged into kumato's repository, but in both cases using different applications, uh we're using engines as a server which are testing against and we are using photo, which is a tool for testing HTTP connections.

B

B

He have found that.

B

In these virtual clusters, our.

B

Solution is quite a bit faster, but before I go through results, I should I would like to remind you that these are virtual clusters with next zero Network latency, so the results aren't really representative of their performance. So this is why I results in the title, but basically well I, have here two examples. First is uh negative: a lot 100 queries per second, with just one thread, and here we see that average mean 50. Percentile is when 12 times lower with our integration.

B

uh 90Th percentile is about two times lower and also we see Max latency much lower.

B

So uh Kumbaya is quite fast. It's just about a few milliseconds of delay, but our integration is faster, but then again it's just a little cluster, so we we're still yet to test it using some production grade clusters.

B

Also, we have I have another example here with infinity Target queries per second, so just how fast it can work- and here we see similar picture I, see that maximum latency with Kuma is the same and I actually think it's not related to Google. It's something, this Linux Network system, performance or maybe kubernetes issues.

B

But overall it's a similar picture. It's about two times faster for 50th percentile, for 90th percentile and about one and a half times faster for uh 99.9 percentile.

B

um I expect that these results will get much less dramatic as soon as we move to production clusters where there will be Network latency, and maybe there will be difference in uh Linux Network stack, but we see that there is some potential people on the performance gains uh So. Currently we are working on setting up testing using production grade clusters like from Google AWS azure.

B

Also previous results: don't really use proper statistical analysis. It's just I read the results manually see that they're kind of the same and do math manually, so we I get to do some statistical analysis and also a better visualization.

B

And so yeah, that's that's kind of it.

B

Thank you for listening and uh yeah. I would very much like to hear some feedback thoughts if it's useful, it fits uh if you have a good setup or maybe I have missed something. Thank you.

A

Thank you, the new uh anyone who has questions.

A

Is the source code to like replicate this test available somewhere.

B

uh We haven't published the scripts to set up it yet because we are still working on it, but yeah we will publish it when we solve for issues with public clusters.

B

A

A

Yeah I have a question uh when you, when you checked Kuma multi-zone, did you use egress or it was just Ingress.

B

uh It was Justin Grace.

A

Okay, so it probably will be worse with egress, because.

C

B

Will be another way, yeah yeah I wanted to because some best case scenario for both uh platforms.

A

Yeah and uh how do you do you have any interpretation explanation of these results like? Why do you think, uh with NSM with squeaker.

B

Yeah uh and Sam actually uses VPP program, which is it translates to Vector packet processing, which is actually even faster than uh Native Linux stack because it processes packets in brand in batches, so yeah, I I, think this should be the main cause of such improvements.

A

A

They do have any other questions.

B

um Yeah I I'd like to add some more questions uh since I have the opportunity about improving our integration so yeah. The first question is.

B

uh Could you think of some real life use cases that uh could be covered by my integration? If we improve it somehow, maybe add something to it?.

A

Yeah I think the use case main use case is when you want multi-cluster and you kind of choose between having Google, multi-zone and or having NSM plus Kuma standalone right. So this is a main use case.

B

uh Yeah but uh maybe there's some issues with an integration that we can solve well, I, guess I. This is not very a good way to ask this question because you probably need to look at the code of the integration to answer it but yeah. Maybe maybe you have something right.

A

Yeah, um to be honest, I like to understand more how like the path of the request like what's what's happening,.

B

Yeah yeah I can discuss it a bit um I, don't think I have any good pictures with me that I can just tackle that I'll talk about it a bit so on each node. We have a hover there and some favor there, which runs VPP and when you start a pod which connects to an SM uh cover the establishes, a connection between it and the pod.

B

And so, if we have a virtual Network, then everything in this network is connected to forward s in on respective nodes, and uh in this veteransi we have a virtual switch which oh, which is also based on VPP, which creates which manages searching between ports.

B

So when you create a pod which is connected to the Fringe, you you get old with additional network interface, for example,.

B

172.0.0.1 and another poet gets another interface with the same network and the network prefixes configured using configuration of NSE Page by the way NC States for network service endpoint. Sorry, so.

A

Essentially, you create like Network overlay, between two clusters and in this overlay you run Kuma Standalone right.

B

Yeah yeah probably disconnect to each other uh with some method, which is also specified in some settings, and so when you uh want to get service from different pod in the same field, three Network, this the packet goes to favor the same node then this for the sends it to uh C. Then it sends to proper.

B

Pods in the first class step also probably for further and the packets are processed using VPP.

A

C

um This might be a hopefully. This is a well-posed question, but like what is, how does NSM compare to other like options for creating and like an overlay, Network just curious, like what the you know, what this I guess, like the selling point or the advantages um yeah.

B

Well, I'm, not a salesperson, but.

C

uh Yeah yeah and that's you know.

B

uh I believe that the main selling point is that uh it is cloud native, so you can deploy some uh puts some deployments into each cluster. uh Then you just add The annotation to pod and to complicated, manages the connection.

C

B

C

Like what's the I guess is a good, um a good example of the of the advantage of NSM, but I. Don't know if you have any other ones, maybe like yeah.

B

Well, yeah, you can. Obviously you can use encryption like with probably any other tools for virtual Networks.

B

There are also options to use avatos, others from VPP, so we can use sriv.

B

If you have any SRI enabled Hardware, then you can here's a special over there to manage it.

B

You can also use uh open policy agent security, configuration I, believe it's called to manage, who can connect to where so, oh, basically, what understand does is it manages VPP and it manages all the management stuff to make it easy to use. I think it's kind of like this.

B

C

A

Yeah, it would be uh interesting to see the numbers from the real clusters when you have like clusters in different availability zones. For example.

A

In that case, uh probably the difference will be that uh dramatic, so yeah yeah expect.

B

The difference will be much less but yeah. It will be interesting to see if it matters. So maybe it's just one percent.

A

Yeah, thank you think of the new effects yeah.

A

We have less than 10 minutes, maybe we should um I, don't know. Maybe we should keep KDs redesign discussion on the next uh I see.

A

A

What do you think Lucas should we discuss this now or maybe next time?

A

If you really want to move to the next time uh depends if you have more questions from community yeah, okay, that makes sense yeah. Maybe then, let's have questions.

A

So, okay yeah anyone has any questions about Kumar.

A

Yeah I also didn't mention that um since the last a community call, we published a few releases. I guess five uh watch releases mostly to update Envoy, and we had some minor issues with disabling IPv6 or for the proxy.

A

And that's pretty much it so yeah new releases are available.

A

If we don't have a any questions, then I think we can wrap this up.

A

Iron: okay, with that yeah, okay, uh have a good day, see you next time right, bye,.

B