The Q&AI Podcast

The Future is Intent-Based: AI That Understands Your Network

• Juniper Networks

Share episode

Host Bob Friday and Juniper Product Marketing Director Jay Gill discuss extending cloud AIOps to the data center using Marvis® Minis and Apstra Data Center Director for proactive application health. They cover agentic AI as the key to automation and the crucial role of networking for AI in maximizing expensive GPU resource efficiency. 

-----

Key points covered:

Cutting through AI hype: Differentiating between marketing claims and the current reality of AI adoption in the data center

AIOps in the data center: Extending the client-to-cloud user experience model using Marvis Minis for application health and proactive maintenance

Apstra Data Center Director vs. data center assurance: Understanding how the on-premises Data Center Director fabric management is complemented by new cloud-based AIOps tools

The rise of agentic AI: Defining the agentic framework as the enabler for conversational networking and true autonomous action in the data center

Networking for AI: Discussing the unique challenges and requirements of high-performance networks in AI data centers to ensure reliable job completion and maximize GPU resource efficiency 

-----

Where to find Jay Gill 

LinkedIn – https://www.linkedin.com/in/jay-gill-0539031/  

Where to find Bob Friday  

LinkedIn – https://www.linkedin.com/in/bobfriday/ 

-----

Keywords

AI for Networking, Data Center, Juniper, Marvis Minis, Cloud AIOps, Agentic AI, Intent-Based Networking, Apstra, Digital Twin, Operations Application Visibility, Self-Driving Networks, Client-to-Cloud User Experience, Data Center Assurance, Congestion Management, Hybrid Cloud, Large Language Models Automation, Troubleshooting 

-----

To stay updated on the latest episodes of The Q&AI Podcast and other exciting content, subscribe to our podcast on the following channels:

Apple Podcasts - https://podcasts.apple.com/us/podcast/the-q-ai-podcast/id1774055892

Spotify - https://open.spotify.com/show/0S1A318OkkstWZROYOn3dU?si=5d2347e0696640c2

YouTube - https://www.youtube.com/playlist?list=PLGvolzhkU_gTogP5IBMfwZ7glLp_Tqp-C

We hope you enjoyed this episode!

SPEAKER_01:

Hello, Bob Friday, and welcome to another episode of QAI. Today we're joined by Jay Gill from Juniper's product marketing team for Data Center. And today we're going to be looking at what does it really mean to extend Cloud IAOps AI for networking to the data center. Welcome, Jay. You know, maybe to start with, give the audience a little bit of your background and where did you how'd you get here?

SPEAKER_00:

Well, thanks, Bob. You know, a long time ago, I was a grad student at Stanford in electrical engineering, and I got really excited about neural networks. At that time, they were pretty primitive. Obviously, the processing power was way lower than we have available today. But I love this idea that you could have this machine that learns and that it then gives you better answers to problems than you could come up with on your own through programming or whatever other means you use. So I've been excited about that idea for a long time. Ever since then, I've been working in technology, mostly in service providers, some in equipment vendors, including Juniper now. And I've always seen how there's really hard problems in networking related to operations, knowing what's going on in the network, troubleshooting, things like that. And the exciting thing is now we can bring AI to bear on that. And I just love that.

SPEAKER_01:

Well, you sound like you're about my age, because you know I did my master's thesis around neural networks in the 80s, also. Yeah, but when did the AI really go from kind of this research topic to something that's really relevant to your job and career?

SPEAKER_00:

Well, honestly, in marketing, I think we were riding the same storm everybody else was when ChatGPT came out a couple of years back, and everybody realized where the technology had come while they weren't really watching large language models and all the capabilities that they bring. And in marketing, that stuff's relevant. So we str we started trying to use it in our job, and we realized some downsides and some limitations, and so we're still adjusting to you know, what can AI do for us and where do we still need to, you know, keep control over the process, have human intuition, have human authorship that really could keep in control of things. But now looking at AI in more specific use cases and networking, I think we're there. We've got some really good stuff to talk about.

SPEAKER_01:

For the product marketing people in our audience, you know, any great examples of where AI has made your job easier or more efficient?

SPEAKER_00:

I think it's been mostly in some of the kinds of things of researching and summarizing, getting sort of decent first drafts of an outline for something together, but not in counting on it for finished products. It's not there yet. You still need an expert human to really curate what you get out of the tools in order to make it into what you really can deliver to customers.

SPEAKER_01:

Okay, now so you know, I always give my market team a hard time about like, you know, let's keep the you know the gap between marketing and engineering reality kind of small. You know, most of this AI marketing looks like a lot of whitewashing.

SPEAKER_00:

Yeah.

SPEAKER_01:

You know, yeah. You're in product marketing, you know, how do you really differentiate AI from all the marketing hype out there?

SPEAKER_00:

Yeah, I think there's there's obviously a lot of hype. And it's something that we in Juniper, under your leadership, are really trying to be honest about and guard against going too far and AI washing. And I feel very strongly about that as well. So in in the data center world, you know, we're honest about where AI is actually being used, where it's making a difference now, and where we're still on the journey to get there. So we fully embrace this journey that Juniper's on as a whole towards self-driving networks. But we are going to also acknowledge that some of that is in the future, some of that's here now, and we'll be honest about telling you where it really is at this point.

SPEAKER_01:

Yeah, I mean, so you know, at Juniper, as you know, myself, right, we've really evolved this AI for networking, client-to-cloud user experience, and kind of really that paradigm shift of not just managing network elements, but really trying to manage this client-to-cloud. You know, how does the data center fit into this Juniper's client-to-cloud user experience story?

SPEAKER_00:

Yeah, I'll say in two ways. First of all, um, the vision of actually trying to measure user experience to simulate that with a digital twin, and then to be able to proactively get rid of problems that users don't even find because they're gone, that can be applied into the data center as a way of thinking about data center troubleshooting and problem avoidance and proactive maintenance and so forth. So part of what we're doing in the data center AI ops is exactly that. Um, Marvus minis are going to allow us to be proactively finding problems before applications see downtime or performance problems. And we've got other tools that are already built into our so-called data center assurance that use machine learning to really do root cause analysis. So there's a lot of ways where we're already getting the benefits of AI in those specific problems. But the other connection point is the data center is going forward for most enterprises, the core of their private cloud. They're going to have a hybrid cloud that includes public cloud instances and private cloud data centers. And if you're looking at managing the end-to-end user experience, you've got to have the experience of the user from the campus and branch all the way into that data center and that application. And so we can connect those things over time and we can really truly measure that end-to-end user experience.

SPEAKER_01:

Oh, okay. Now you mentioned a couple of things here. You know, Marvus Minis, you know, as everyone knows, you know, campus and branch introduced Marvus Minis last year as kind of this digital twin to really make sure that critical network services and applications are actually up and running in the campus branch, you know, before the business opens. How does Marvus Minis relate to the data center?

SPEAKER_00:

Yeah, it's um it's an analogous situation. You know, in the data center, you're not necessarily right next to the user and thinking about measuring their experience in a branch office, for example, but you're right next to the application. And so you think about what's the experience of that application. Is it is the network available for it? Is the network performing? Can it reach all the critical services that it needs, like a DNS server and so forth? So we can emulate that thinking about really simulating the the experience of the application and having the digital twin of the application to see am I experiencing any problems? Will an application experience problems when you try to spin it up, you know, today or next week? And so I think we're learning from that experience, uh, Marvus Minis in the campus and branch, and just trying to take the best parts of that into what we can do in the data center.

SPEAKER_01:

Okay, so the same thing as the campus branch, we have DHD, DNS, all these services have to really be up and running for the data center also.

SPEAKER_00:

Yeah.

SPEAKER_01:

Now, client to cloud, you know, to your point, whether it's public cloud or private cloud, we have applications running in all these different things. Right. You know, from your marketing experience, any good stories where you see customers really interested in making sure that, hey, if they have a critical application, you know, and they're trying to find out, you know, is it the campus branch or the data center? Any great stories where you see customers really want to tie their branch and data center experiences together?

SPEAKER_00:

Absolutely. So we have a large entertainment customer who's a customer of both our MIST wireless and campus of branch solutions and our data center solutions. And when we were able to turn on the integrated, you know, Marvus Actions dashboard, which shows the whole network from one end to the other, from the from the branch all the way into the data center, and they could have that cup of coffee view that showed anomalies across the whole network, they loved that because now they realized that if there was a problem affecting any part of that user experience, they were going to see it on that dashboard and then be able to start drilling down. Same thing actually applies for Juniper IT. There are users of our products both in the campus and branch and in the data center, and they had that same reaction.

SPEAKER_01:

Yeah, somebody, you know, being the co-founder of NIST, you know, Juniper, we have this day zero, day one, NIST Marvus, you know, helping customers get their networks up and configured. I know in the data center we have Apstra, who acquired in, I think, 2021 or so. You know, maybe give the audience this little background, you know, what is Appstra?

SPEAKER_00:

Yeah. So Apstra, you're right, has was founded as an independent company, acquired into Juniper a few years back, and it still goes strong as the core of our data center fabric management and automation platform. So what Appstra does is it's an intent-based networking engine. That means that you tell it what your intent is for this network, how it should operate, and it actually configures all the switches and devices in the network in order to realize your intent, and then it monitors it continuously to make sure it doesn't deviate from that intent. So that's a deterministic control model of making sure your network is exactly the way you want it. But stuff happens in the network, and so you have to be able to also respond to unpredicted failures, to other things that might be happening that throw that network out of whack. And that's where day two assurance becomes a big deal. Abstra does some of that, but AIOps and cloud-based AI tools can do even more of that. And that's where there's a complement between Apstra running the network, managing the fabric, and AI ops to help you with those tough day two troubleshooting and resolution problems.

SPEAKER_01:

So I know that we have kind of this missed Marvus going on for campus and branch to help with day zero, day one, day two operations. Yeah. I think in the data center business, I saw, you know, we just announced Astra Cloud, Data Center Assurance, you know, and I know we're trying to extend Cloud AOPs across both campus branch and data center.

SPEAKER_00:

Yeah.

SPEAKER_01:

You know, maybe contrast Apstra versus Data Center Assurance in the UK.

SPEAKER_00:

They're complementary tools. You have Astra working on premises to manage that day-to-day operation configuration of the data center network. And then you have the cloud-based tools, you know, building on the platform that you built with MIST and the MIST team built, taking those tools and all those capabilities and all that learning and applying it to data center. So we started out by building new applications for data center insurance in the cloud that do a few things. They bring in application data and map it to the network so you can really see what applications are running on the network, where the flows are, where there might be imbalances in traffic, where what would be the impact if this switch went down on applications. Usually that stuff is opaque to the data center network operator. We're giving them visibility into the applications running on their network, and more importantly, the ability to drill right to the root of the problem because the AI tools have already looked at alerts and failures and alarms and zeroed in on the root cause. And so you go straight to the problem you need to solve. So that set of tools, all those AI capabilities in the cloud, complement what Appstra was already doing on premises. And the other way that they complement is you always say, and we always say, good AI starts with good data. And in the data center, Appstra is the source of that good data. It's got the entire network essentially stored as a digital twin. It's monitoring telemetry data and updating the state of the network. So you can query it at any given moment and say, what's going on in my network? What's the state of the network? And that's the data that we build AI applications on.

SPEAKER_01:

Okay, yeah. Now, I can't help it, you know, things are changing so fast in this industry, right? AI, exponentially changing. You know, AI is actually kind of passe now, right? It's like as so yesterday. Right. We're in the agentic AI now, so we can't really have a market session without talking about agentic AI. Yeah. So, agentic AI, AI, what's the difference between agentic AI and AI from last year?

SPEAKER_00:

Well, I think people are still coming to grips with agentic AI and coming to common understanding of what it is. And we're not quite there yet. But what I would say about agentic AI in the data center, the way we see it, is that it's a framework for allowing AI to interact more specifically with tools like Abstra. So in the data center, what we're doing with Agentic AI to start out is building the framework that allows things like the Marvus Conversational Assistant to talk directly to the network. And that means the network operator can ask questions in natural language and get data from the network. And the next step will be allowing that network operator to just tell Marvus in natural language what needs to change in the network. Add a VLAN. Show me all the switches that are at more than 50% capacity, update this configuration. Whatever it might be, more and more the operator will be able to interact with the network through the conversational interface and through Abstra because we have that agentique framework there. And then it learns and it becomes better at even taking autonomous action as long as the humans have learned that it's trustworthy.

SPEAKER_01:

Okay, so you're a marketing guy. So I usually tell people is hey, agentique eye is really a new, non-deterministic, non-linear software programming paradigm, right?

SPEAKER_00:

I like that, yeah.

SPEAKER_01:

Something that's you know it's really a catalyst to the next step in the evolution of automation. You know, so when you're talking to customers and they're trying to tell the difference, you know, is this a Gentic AI versus AI? You know, any good use cases that would really resonate with, you know, hey, if I want to know if there's a Gentic AI, what should I be seeing?

SPEAKER_00:

Yeah, it's a good question. I think what we're seeing a Gentic AI solve first is just giving people a speedier way to get answers to questions. There's tools now that you can go into a UI point and click and use menus and so forth to get the data you want. But if you can just ask that question and have the agent tell you the answer, because it's smart enough to learn both what you mean by the question as well as what data is available and how do I get it, and how do I put it together to make a sensible answer. That's really speeding up just everyday operations. Almost anything you can think of gets sped up when once you put that framework in place. And as it learns over time, it'll start to know that any kind of anomaly or problem that's out there, maybe it already knows what action it took because the operator told it to last time. Maybe it can learn to go ahead and take that action and the users can trust it. So that's where we're going, towards self-driving. And the agentic framework just sets us up to get there.

SPEAKER_01:

Okay, so now maybe to wrap up this episode, you know, we've been talking about AI for networking. You know, you hear Rami talking about the other side of the coin, networking for AI.

SPEAKER_00:

Yeah.

SPEAKER_01:

You know, can't really leave the episode without a little bit of networking for A. What does that mean? Bread and butter for data center?

SPEAKER_00:

Absolutely. So AI data centers, the places where you're doing model training in order to build these models that we use, and also inference to use them and give results back to the people who are asking questions of the AI agent. Those data centers are massive compute resources, GPUs and other kinds of compute resources, connected by massive switches. And so they're just big special purpose data centers. There's a lot of money going into building these things. You know, hyperscalers spending huge amounts, but a lot of other players are building their own capability as well. And so the question becomes: what kind of network do you need for that AI data center? What makes it special, different from other data centers, and what are some of the key problems you're trying to solve there? And I'll just boil it down to one thing, which is reliable job completion. If your AI model training is running and GPUs are talking to each other across this high performance network, and there's a network hiccup, there's a packet loss, there's high latency, those GPUs have to stop and maybe even restart back to where they were three hours ago. And you've just wasted a huge amount of very expensive, very valuable computing capacity when that happens. So your network can be the source of a lot of wasted money. If you do it right, the network becomes the enabler to maximize the use of that very expensive resource and make sure that you're training your model with the lowest job completion time and the lowest overall cost. And you've got to be able to manage a variety of things that are unique to AI data centers when you do that as well. So Appster comes into that equation as the manager of an AI data center, and some unique capabilities like congestion management have to be built in. So we're we're solving all those problems in order to build those data centers that enable all the AI applications we all love to use.

SPEAKER_01:

Okay, well, Jay, I want to thank you for joining me. It is clear that AI is going to be a transformational society. These data centers are gonna be at the foundation of that transformation. So thank you for joining us. Thank you for joining us today at the episode, and look forward to seeing you at the next episode of QAI.

Head Shot

Bob Friday

Host
Head Shot

Bob Laliberte

Co-host
Head Shot

Mike Spanbauer

Co-host

Podcasts we love

Check out these other fine podcasts recommended by us, not an algorithm.

Packet Protector Artwork

Packet Protector

Packet Pushers