Cloud Native Computing Foundation ServiceMeshCon EU 2022, 19 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Lightning Talk: MeshMark: Service Mesh Value Measurement - Lee Calcote, Layer5 & Mrittika Ganguli

Description

Lightning Talk: MeshMark: Service Mesh Value Measurement - Lee Calcote, Layer5 & Mrittika Ganguli, Intel

Still trying to understand how to best gauge the performance of your cloud native infrastructure? Confused as to whether self-published, performance benchmarks are trustworthy or simply biased marketing in disguise? Measurement data may not provide a clear and simple picture of how well those applications are performing from a business point of view, a characteristic desired in metrics that are used as key performance indicators. Behold MeshMark: a performance index that provides you with the ability to weigh the value vs overhead of your cloud native environment. Convert performance measurements into insights about the value of individual, cloud native application networking functions. Join us as we distill a variety of microarchitecture performance signals and application key performance indicators into a simple scale. Explore the other side of the performance measurement coin: value measurement.

A

Good morning, good afternoon we are presenting a cloud native benchmark index. Today I am ritikal ganguly from intel, and my partner in crime is lee calcott from layer 5.. This is a cloud native value measurement and it's pretty exciting. So let's get started.

B

We're, uh as you know, we're missing some performance characteristics.

B

um We are uh well well, any number of you in the audience have a lot of metrics that you use to track your environments um and actually maybe too many it might. Actually, if you think about it, might actually take you a little while to articulate the and characterize the performance of your environment.

B

If you had to boil that down to a single metric, maybe to a handful of metrics, what would those be golden signals? Maybe some signals from the red method or if you had to boil it down to one. You know I don't know that we have a just a clear and concise way of conveying.

B

You know the of characterizing, the performance of your environment.

B

Nor do we have a single kind of true north uh measurement, we're also, you know quite frequently overlooking business performance, which is in large respect to why we're running the infrastructure. In the first place, we usually talk about performance in cold, hard, quantitative speeds and feeds, but instead I would submit to you that performance could should be. You know it should absolutely be measured in terms of speeding, speeds and feeds, but it's a lot more meaningful to layer in the value and to quantify the value that your infrastructure is providing.

B

So we're really kind of missing the business performance aspects of what we're tracking, how we're characterizing. So the discussion that we're having today falls under the umbrella of a cncf project, the service mesh performance project.

B

This project is, at its core, probably a specification for capturing the details of your environment in a uniform way in a consistent way, capturing your infrastructure, configuration your service, mesh configuration and characterizing the details of your workloads and doing so consistently such that you can baseline your environments. You can benchmark them in a consistent way, share with others. Maybe compare with the performance that others are having to the extent that it's codified, you can have system to system exchange of this information, and so smp is a cncf project as of well.

B

I guess it's maybe less than a year now. There's some research aspects to what service mesh performance is as well.

A

Sure, and so we have all been involved with the uh with s p project for more than a year now, with people in our teams contributing and it's an extremely useful.

A

You know set of tools and as part of that we introduce meshmark mark, uh like I mentioned it's a cloud native value measurement uh from value you're, essentially trying to measure if the performance of your infrastructure matches what you want to get from your deployment. What kind of value you want to get business value you want to get from your deployment.

A

So, for example, if you have some key performance indicators, do you want to measure whether the mesh mark value is?

A

You know directly responsible for how say your video gets uh loaded or your image gets loaded on a particular web page.

A

So often you'll see when you have a youtube video uploaded, it will have only the text and not the video, and so uh what happens is you do not want that kind of an experience you want the video to be and the images to be loaded first, as you see in this video here inside etsy, if you click on something you may often see the text get, you know rendered first and then the video, the load latency of the video traffic is what impacts uh you see visually, and so the deployment of your cloud native environment is now.

A

If it can be indexed through a mesh mark ratio, your load latency will be directly proportional to that, and then the number of resources you are using to deploy this environment to get a particular lower, latency or a higher latency, is a usage metric and that's directly proportional to your mesh mark. The tco of your application hence becomes directly related to meshmark, and so how do we go ahead and define what is mesh mark? And here is a definition that we propose and we are working towards uh building this up.

A

Essentially, you have utilization of different kinds of resources, and these utilization classes are can be categorized as, for example, a compute utilization or a network utilization or any other resource utilization could be a utilization class and, as part of that class, you will have some efficiency metrics, which we call mue and we'll define that attach weights to those mue have a set of 1 to n metrics with weights, and you sum them up as a ratio of the number of mues and that becomes your mesh mark, you can give weights, you can give higher priority to utilization, 1 or n, and you have mesh mark which and one index which can tell you how well it is mapped to your value vector.

A

So as next steps, we look at one specific mue. So what is an mve? It's a calculation, combined uh ratio of measured platform resources to assign resources.

A

If you are able to um measure what your assigned resources are in whatever form and able to also uh you know monitor, what's the measure of what's the used resources, you can have this ratio so, for example, a very simple one is cpu performance and you would want to see if the cpu performance, uh as uh as a as a ratio to the available resources, is a loss or a gain so cpu performance raw loss over total cpu is our mue1 and that's just one minus cpu utilization over 100, that's a very simple ratio, and if you see on on the right hand side, the graph shows you that, as the latency increases your mue lowers, and so that's a very good indicator that your efficiency of your infrastructure is not very good because your latencies are increasing as your qps increases.

A

So it's very easy to then take action based on my emu is lower. Let me take action and that's something we can address in future discussions. So like this, you can, you know, measure and create other mues.

A

We will look at how to uh how you can visualize this within an environment, and so, let's look at the demoli.

B

Sure, let's jump into a sibling, cncf project called meshri mesheri is a cloud native management plane. Users of meshri can configure their kubernetes deployments any and every service mesh, as well as on board and off board their workloads onto any given mesh.

B

Let's take an example: workload, a console application load it into the visual designer, take a look at the service, splitting functionality of console and note, in this case we're assigning a weight of three when we can change that to four to derive its mesh mark, which is a mesh utilization efficiency calculation of the efficiency by which that network function is being performed. We could also take a look at service intentions of console and examine the efficiency of that network function.

A

Now that you've seen the demo if.

B

A

Want to go ahead and publish the results and call everyone to get together, give us feedback, and if you have a unique use case, join us. um What's say what else.

B

Yeah, we are continually looking for people to give descriptions of their workloads. We want to make sure that the workloads under test are, you know fairly. Accurate representations are interesting representations to you.

B

There are a few different predefined tests that are being run and those are being run by some of the contributors in the project within the cncf labs, so within controlled environments.

B

At the moment there are upwards of about 40 000, test results that have been collected, and so a lot of data to be analyzed and to be published. So part of the goal here is to begin publishing. Some mesh marks ritika thanks so much for engaging in the definition of meshmark. This has been amazing.