Microsoft Research Podcast - 085 - Live video analytics and research as Test Cricket with Dr. Ganesh

Episode Date: August 14, 2019

In an era of unprecedented advances in AI and machine learning, current gen systems and networks are being challenged by an unprecedented level of complexity and cost. Fortunately, Dr. Ganesh Ananthanarayanan, a researcher in the Mobility and Networking group at MSR, is up for a challenge. And, it seems, the more computationally intractable the better! A prolific researcher who’s interested in all aspects of systems and networking, he’s on a particular quest to extract value from live video feeds and develop “killer apps” that will have a practical impact on the world. Today, Dr. Ananthanarayanan tells us all about Video Analytics for Vision Zero (an award-winning “killer app” that aims to reduce traffic-related fatalities to zero), gives us a wide-angle view of his work in geo-distributed data analytics and client-cloud networking, and explains how the duration and difficulty of a Test Cricket match provides an invaluable lesson for success in life and research. https://www.microsoft.com/research

When we started, we looked at the winning entry in the tracking challenge, which is essentially a challenge where the greatest computer vision people compute to produce the best object tracker. And what we saw was that the winning tracker runs at one frame a second on an eight core machine in parallel. And to just give a perspective, a camera's frame rate's at 30 frames a second. So it was kind of clear to us at that point that if video analytics was to take off, systems and networking folks needed to put their heads into the game and jointly work with the vision researchers to get a working solution.

You're listening to the Microsoft Research Podcast, a show that brings you closer to the cutting edge of technology research and the scientists behind it. I'm your host, Gretchen Huizenga. In an era of unprecedented advances in AI and machine learning, current-gen systems and networks are being challenged by an unprecedented level of complexity and cost. Fortunately, Dr. Ganesh Ananthanarayanan, a researcher in the Mobility and Networking Group at MSR, is up for a challenge. And it seems, the more computationally intractable, the better. A prolific researcher who's interested in all aspects of systems and networking, he's on a particular quest to extract value from live video feeds

and develop killer apps that will have a practical impact on the world. Today, Dr. Ananthanarayanan tells us all about video analytics for Vision Zero, an award-winning killer app that aims to reduce traffic-related fatalities to zero, gives us a wide-angle view of his work in geo-distributed data analytics

and client cloud networking, and explains how the duration and difficulty of a test cricket match provides an invaluable lesson for success in life and research. That and much more on this episode of the Microsoft Research Podcast. . Ganesh Ananthanarayanan, welcome to the podcast. Ganesh Ananthanarayanan Thank you. Thanks for having me here. Laila Lalami I've heard some people say your last name

could be your secure password. Ganesh Ananthanarayanan I spelled it wrong once. The first year in the US when I had to file my tax return, I added an extra A.N. in the end. And so as a result, the IRS never sent me my refund. That's awful. So it wouldn't be a useful password if I get it wrong. Well, and you told me something interesting about what your name means. Yeah, the name actually translates to something like something that never ends. It's unending, and it's kind of

apt, isn't it? Yeah. Well, I've got a crazy last name too, but as I said, welcome. So you work in mobility and networking here at Microsoft Research, and your work is broadly situated in the area of systems and networking. So we're going to do a deeper dive into your current areas of research shortly, but to kick us off, tell us in general what gets you up in the morning. Why do we need systems and networking research and why are you doing it? That's a good question. I mean, what gets me up sort of like professionally in the morning is frankly just the job I have. I just love this job description I have where I can sort of like creatively explore the topic that I am interested in,

keep an eye out for all the trends that are out there, and at the same time, you know, work with this huge sprawling infrastructure that a company like Microsoft has where it can affect millions of users. So to me, that's like the best job description, and that's frankly what gets me excited every day. Systems and networking specifically, I like it because from many, you know, sort of like even a very philosophical level, I connect to it because the core aspects of systems and networking are, you know, there are trade-offs all the time. There's no single

perfect solution, which is kind of true if you think about it in life in general. And the second thing is, you know, you cannot hide behind idealized models. Success is what works out in the real world. And that's kind of true with, you know, life in general and systems and networking. So in many ways, I connect with it. Well, you know, we're going to go sort of all over here because you have such a broad range of interest. But you have a big interest in live video analytics, which we'll get to in a second. But I want to stay upstream for a minute to set the stage for that.

And one of the things that shows up in the talks you give is this phrase, cameras are everywhere. It's both reassuring and frightening at the same time. So we'll get to frightening later in the podcast. But for now, I want you to answer two questions. Why are there cameras everywhere? And why is this a good thing? In many ways, I think the falling camera prices contribute to it. But the main reason I think cameras are seeing this pervasive deployment is the fact that there is this ability for you to monitor things, to record things. And many ways, I think people put these cameras in there in sort of like the early stages with the hope that there would be a human who is going to look at it and perhaps react

to certain things and so forth. And as we saw this potential with what, you know, computer vision could do that, hey, I could actually analyze these cameras in real time, perhaps if needed, these camera streams, and do a bunch of good stuff with it. So that was kind of what we went with in terms of saying, hey, what are the kind of things that we can enable with these cameras that frankly can be very beneficial to us as a whole? I can give an example of a recent demo that we did at the Microsoft TechFest.

We demoed a smart crosswalk. But if, say, somebody with special needs comes to the edge of the crosswalk, say in a wheelchair, then we would detect that automatically and then have the walk light turned on for them. And not only that, what we also do is that as the person is crossing, if you notice that, hey, the timer is going to go off, but that this person at the pace at which they're walking needs some additional time to finish crossing safely, we give like an additional, say, 10 seconds or so, so that they can finish crossing. So I see a lot of potential for a lot of good that can come out of just doing these video

analytics, and that can happen only if cameras are everywhere. I'm thinking of all the places that I know there are cameras, like convenience stores, traffic light cameras, traffic cameras in general on the highways, retail scenarios. And then we go back to this idea that somebody should be able to take those images and do something, usually after the fact, right? When did it become so overwhelming that we needed computers in the mix? So in both cases, I think even for after the fact, if you have to search through months of videos, there is like a rate at which you can process it if a human is viewing it. Maybe we can view it at 2x the speed, but that still would take a lot of time if I'm viewing two months worth of video. And then, of course, then there's additional

point where, hey, why should we stop at only doing things after the fact? Like the crosswalk scenario I just explained, that can happen only if we analyze the feeds in real time. So that's when I think we started getting into this thing where, hey, we needed to actually analyze the feeds in real time. So that's when I think we started getting into this thing where, hey, we needed to actually analyze these feeds in real time, live video analytics. MELANIE WARRICK- Talk about the challenges that you faced once you had the aha moment to be able to technically accomplish that.

SHYAM GOLLAKOTA- So live video analytics is very challenging from a systems and networking perspective, as we see it. Among the various domains for, say, systems built for AI, very challenging from a systems and networking perspective as we see it is you know among the various domains for say systems built for AI this was the most challenging because you know the amount of compute that is needed the amount of data that is generated the amount of storage that is needed the kind of privacy concerns that come out of videos. These are all nothing like any other domain. And if you look at, you know, what I listed in terms of compute, the network, the storage, the security

and privacy, these are all the major pillars of systems and networking and video analytics stresses each one of them. So from that point of view, we thought it was a pretty, you know, interesting challenge to take on. And then what we also noticed was that, you view, we thought it was a pretty interesting challenge to take on. And then what we also noticed was that, obviously, we got excited to doing video analytics because we saw the possibilities of what computer vision could achieve. But when we started looking at the actual solutions, we also noticed that, hey, many of these solutions that are kind of cool require an enormous amount of resources for it to run that it's almost impractical

if you were to run something like this 24-7. Just to give a concrete example, when we started, we looked at the winning entry in the tracking challenge, which is essentially a challenge where the greatest computer vision people compute to produce the best object tracker. And what we saw was that

the winning tracker runs at one frame a second on an eight-core machine in parallel. This is for a single camera. And to just give a perspective, a camera's frame rates are 30 frames a second. So it was kind of clear to us at that point that if video analytics was to take off, systems and networking folks needed to put their heads into the game and jointly work with the vision researchers to get a working solution. Okay, so let's say you've linked arms now. What is your quest? What are you trying to accomplish in this field? So what we are primarily out to do is to democratize video analytics, which is sort of being able to analyze videos live in real time at low cost, and at the same time produce results that are accurate enough for the task at hand. How's that going? It's going pretty good. We've been at it for a couple of years now. We've had many systemic breakthroughs in terms of how we could make things substantially cheaper. We've engaged with a bunch of customers in terms of analyzing their needs and so forth.

So I think we have like a good sense for what it is that people want with video analytics. So I would say we're quite excited in terms of how it's progressed so far and the road ahead. All right. Well, we're going to get on that road pretty quick here. It does make me laugh, however, to note that what you're aiming for is cheap, fast, and good.

We're all looking for that holy trinity of cheap, fast, and good. Usually they say pick two, right? If you want it cheap and fast, it's not going to be good. If you want it cheap and good, it's not going to be fast. And video analytics is like the Uber challenge.

That's actually true. I quote this point actually in my talks too, that in this goal of all the three axes, I usually say that if you think about it, achieving any two of the three is much easier than aiming for all three. So I totally agree with you. So you've given yourself like the hardest task. Yeah, video is the hardest.

I mean, we're also looking at it, and if you look at, say, the whole space of Internet of Things, right? Yeah. Cameras are the hardest of things. So, yeah, I think if we get this right, then we should be able to, you know, navigate through a whole bunch of other verticals as well in this space. Right. That's exciting.

Well, let's get specific. And I want to focus this part of our conversation around your work in healthcare. Okay, I'm making air quotes around that. Okay. Because you're not doing health care, but, and I didn't know this before I talked to you, traffic accidents are among the top 10 leading causes of death in the world. And all the others that surround it are actually health care issues. So you've got a hero app for that.

You call it Video Analytics for Vision Zero. Tell us about it. Yes, you've got a hero app for that. You call it Video Analytics for Vision Zero. Tell us about it. Yes, you're right. It kind of shocked us as well to hear that traffic-related fatalities were among the top 10 reasons for deaths worldwide. So, when we started out this project on video analytics, we wanted to really have like a driving application where this was really useful. So at that point, we reached out to the city of Bellevue, where we wanted to check if we could utilize their traffic cameras for any kind of analysis that they would want to do. And it had so happened that the city of Bellevue had signed on to this program called Vision Zero. Vision Zero was a larger initiative that started in the late 90s in Sweden, where the idea was to get the number of traffic-related fatalities to zero. And so the

city of Bellevue was interested in partnering with us, and we had like a very forward-looking and enthusiastic partner in France, Lohenherz, in the city of Bellevue. And along with France, the question that we asked was, how can we use these cameras to improve traffic safety, traffic efficiency, and just long-term planning of traffic? So we worked with the city of Bellevue in terms of analyzing their live feeds to produce, you know, live traffic counts to raise alerts when the amount of traffic is abnormally high or low. We've helped them in terms of a bike planning initiative that they had, where we wanted to do like a before-after study on how the number of cars and the number of bicycles changed.

So overall, this has been an initiative that in many ways has been sort of like trendsetting. We've presented this work at a lot of transportation engineering forums, and many cities have expressed interest in terms of adopting a solution like this. So in many ways, we believe we've opened the eyes of folks to what is possible with video analytics in the space of traffic cameras. So that's been pretty good. Let's get a little more technical here. Since we're talking about cars, let's look under the hood of video analytics for Vision Zero and talk about Rocket,

which is the aptly named video analytics engine behind it. So break down the technical components of Rocket and explain how this approach is bringing some of those traditional barriers to accurate and affordable video analytics to real people? Rocket, I think we had like a poster title once that said Rocket to enable video analytics to take off. So the Rocket stack has many components and right at the top of the stack is this ability to express video analytics as a pipeline of various operators. So we

needed those abstractions where somebody can look at it and go hey what is the set of operations that I should string together for me to run this video analytics application. Specifically an example of that could be take the video stream, run a decoder, maybe run a filter where some of the frames can be filtered off, then you run an object detector and say then a classifier. And then, of course, in the end, you have some custom logic, say, for counting and so forth.

So this is what sits right at the top of the RocketStack, this pipeline expression. And then we have the pipeline optimizer, which is kind of important because what I explained right now is an example of a generic pipeline. Different choices have to be made for different pipelines, depending on what the camera stream looks like. Right. What is the kind of filter that I should use that's best for this specific camera stream? What's the kind of object detector that suits the best?

What's the frame rate at which I could be running it? Like, for instance, if this is a stream where, say, it's at night, there's not that much activity happening, I can sample out a lot of frames and I can save on resources. Whereas if it's in the daytime, I don't want to be sampling frames off too much. So this is kind of what the optimizer makes a decision on, which is what kind of configurations should we be picking for this specific camera.

And then after this comes the resource manager, where it looks at not just one specific pipeline, but say a bunch of pipelines together. Presumably a city has hundreds of cameras and all of them are sort of like running together in a cluster, so to speak. So if I have these resources, how is it that I would distribute these resources

across these different pipelines? Clearly, I want to be giving resources to those pipelines that need the most to get the sufficient amount of accuracy that they need. So this is kind of what the resource manager does. And once this is done, we move to an important portion of the project where video analytics intersects with edge computing, where we are not going to be doing video analytics,

especially at scale, if all the videos have to come to a single central place in the cloud. So video analytics has to be split across a bunch of smart cameras, some edge clusters in between, and then the cloud. So after the resource manager does its thing, we move on to this edge cloud executor that splits these video analytics pipelines between the various edge clusters and the cloud.

And finally what we do is, hey, we are running live video analytics already. Can we use the results from this to just index the times when, say, a car showed up and an accident happened or a near accident happened? Store them so that later on, if you ask me a query on, say, in the last month or the last two months, can you find me the times when there were accidents?

You don't have to process the entire set of videos at the time you're asking the query. You should just be able to pull it off and index. So this way we can actually answer queries on even stored videos, like almost like three orders of magnitude faster than otherwise. All right. So you've partnered with, say, the city of Bellevue, and it's actually, was it a research project or have you implemented this or deployed this anytime, anyplace? We've actually implemented, deployed it and had it running for a long time. So we went through this entire thing where we built the system. We produced, say, live counts, alerts, and a whole bunch of stats for the traffic. We built a dashboard that displayed all these counts live in real time, including alerts. And then this dashboard running live in the city of Bellevue's traffic management center.

And so, yeah, we wanted to do this entire thing where we went beyond just, you know, saying, hey, this is the part that I am interested in. Instead, looking at it holistically, like what does it take to build the end-to-end product, if you would, for something like this? Well, right about now, I want to tell our listeners that they ought to go to the website and look under your name.

There are talks, there are videos, there are projects and papers all about this. It's a very visual thing that you're doing. And there's a lot of video on there that people can see this in action. It's really cool. I looked at it. In fact, when I was looking at the Vision Zero in Bellevue, I was actually looking for my own car had they caught me. Yeah, like one of the cool things to do sometimes when I drive around with friends is, so one of the cameras we analyze is right next to Lincoln Square in Bellevue. And it's cool to point them to that camera and say, hey, that's one of the cameras that we are analyzing.

So smile. So smile. Selfie. We're just on camera, yeah. All right. Well, let's talk about some of the other cool work you're doing. One of the attributes of next-gen data analytics is that workloads are no longer confined to one data center, but spread out across many data centers and edge clusters around the world.

And this poses a host of challenges for systems and networking researchers. So tell us about the challenges and how you're tackling them through your work on geo-distributed data analytics. So my background during my PhD centered a lot on big data analytics. We did some of the earliest work in terms of straggler mitigation in big data analytics, which is kind of important to, you know, maintain a whole bunch of service level objectives for these big data jobs. And after that, we're starting to think like, you know, what is like the next generation of data analytics going to look like? And we saw that, hey, these companies like Microsoft, you know, the Azure infrastructure is sprawling. It's not just a single data center,

but spread out worldwide. And these are not just data centers. We have a whole bunch of other edge clusters and so forth. And in many ways, you know, what we figured was that, hey, analyzing all this data that is sort of like geo-distributed now is the next sort of like frontier, if you would, for big data analytics. And so what we sort of like ended up with was two key results or accomplishments. First was that we needed to do these data analytics without necessarily aggregating all the data to a single place. Because if we were to aggregate all the data to a single place, the network would be too much of a bottleneck, and that's not the way we should be doing it. And the second thing we wanted to do was to make sure that queries and the whole infrastructure that we ran in a single data center

should automatically cross over when you run it across a cluster of data centers as well. So we had this abstraction of sort of like an internet-spanning cluster, again, of all these different data centers and edge clusters worldwide. And so then we designed a bunch of solutions that essentially dealt with the enormous heterogeneity that you would see in these different clusters, the different network uplinks and downlinks. But the cool thing was that

you could write the same query and it would run both in a single data center as well as distributed across many data centers and edge clusters. And this in many ways, this work, the way we're thinking about it was what's also, you know, to me personally led to this topic of video analytics because I was thinking more and more about it and it started becoming clear to me that one of the biggest sources of distributed data that's generated is going to be cameras. How is it that you're going to analyze all these camera streams without bringing it all to a single data center? All right, moving on. My life is literally dependent on the rather tenuous idea that all my technologies are going to get along with each other all the time. Of course, everything doesn't always go perfectly.

Things go wrong. And that's why there's people like you to work on the systems and networking problems that we face. And as we depend on our networked technologies and increasingly mission critical world. Sure. So I focus on this wide area part of the network or more specifically the part of the work that's happened almost like in the last decade in terms of the networking inside a data center and connecting our own data centers that the client cloud part is the weakest link. And that's kind of what we really need to address and take care of. Okay. And it's kind of interesting. you also brought up this thing about, you know, whom to call, because one of the projects that we are actually working with right now with Azure networking is this problem where when connectivity is bad between our clients to any of our data centers, who is the culprit? Whom do we blame?

We actually call the project blame it. So if you think about it between the client and us there are a bunch of autonomous systems or ASS in between and the Internet is designed for these autonomous systems to function at on mostly so it's kind of an interesting challenging problem in terms of how is it that we'd zone in without necessarily owning all the pieces between us and the client. So this is work that we've been doing with Azure networking. We've deployed initial parts of it. And so, yeah, we're excited about where it can progress.

Okay, so without sort of signing an NDA, tell me what you're doing. Technically, how are you? Are you employing machine learning and AI techniques to, because this is something no human, not even a bunch of humans that have brains like yours could do. So in fact, it was interesting since we brought up machine learning is the way we got interested in this was that a couple of years back, we were working with the Skype team in terms of relaying Skype calls. Say the caller, say, is in Seattle. Theee is somewhere in India, in Bangalore. They could get into the relays, say in Bellevue, go through the Microsoft backbone, and then exit at some point close to the callee.

And so it was sort of like a problem about which parts to choose. In many ways, we mapped this to a multi-armed bandit formulation. Interesting. Where we could sort of like explore all the parts that are out there for their different performance characteristics, and then exploit the one that is the best and continuously do it. Well, while we're on that subject of machine learning and systems and networking, let's talk a little bit about your split personality. At least part of you is a networking researcher, but the other part of you is interested in

machine learning techniques and so on. So how do you bring yourselves together and how does this inform the future of networking research? I primarily see myself as a systems and networking person. That's kind of what my training has been, a bunch of my work has been in as well. So what I do is bring the systems and networking lens to a bunch of other problems, like specifically these days with vision and ML, where how is it that you can make something more accessible and easy?

How is it that you can make something more cheap? How is it that you can make something more expressible and easy? How is it that you can make something more cheap? How is it that you could make it more reliable? So that helps where we could get like this bunch of trade-offs that I had mentioned earlier in our conversation for video analytics, for instance. And so that's the direction I come from. But there is also this other direction, which is often kind of useful too, which is a whole lot of problems that are there in the systems and networking space could actually benefit if you can take a look at it from an AI lens or an ML lens. Right.

I had mentioned about the bandit formulation, and this is a sort of like a longstanding problem, but how is it that you choose the best path between two hosts on the internet? I remember we had a panel at the Microsoft Faculty Summit last year where we hosted a bunch of people on this topic. I think if I recall right, the topic of the panel was the good, bad, and the ugly of ML in network systems.

So we were debating about the set of topics where ML is appropriate in a networking context. And I recall one of the points that we had discussed was that it's good to have it in cases where it's computationally intractable, like, say, scheduling or query planning. And I remember we also had an Azure person on the panel. And his point was also that we also need to get to a point where the results are explainable, because if something

breaks in production, the last thing I want to hear is that, hey, I don't know why this happened. So yeah, but a bunch of people have already shown the potential of applying ML to systems. But I think the journey ahead and the possibilities ahead is quite exciting of what we can apply and do. All right. Well, we've reached the point in the podcast where I ask what could possibly go wrong. And I think we have to go back to the phrase cameras are everywhere. We've talked about the upside and how you're working on applications that help people. And I think we can all see the benefits there. But as we know, with every powerful technology, there's always a downside.

And some people would argue that you're creating, or if not creating, equipping Big Brother. So is there anything about your work that keeps you up at night, Ganesh? This definitely keeps me up at night in that, hey, by making video analytics cheaper, is it that we're enabling more and more people to be surveilled. So I look at this from the paradigm of mechanism and policy, where I see what I'm doing as mechanisms for doing something. And then there is a whole set of policies that sort of like explain and stipulate what can be done and what cannot be done. And so it's heartening to me that this debate around policy

has in many ways already begun. People are already taking it quite seriously. You know, like for instance, we tried to work with the Microsoft surveillance system. And while on one hand, it was frustrating, we could not get access to those feeds. On the other hand, it made me feel good about the fact that, hey, somebody is taking it very seriously that even with Microsoft employees, these are not videos that would just be allowed to hang around there. LESLIE KENDRICK MASON, JR.: Be data for research. SETH VARGO, MD, PhD.: Yeah.

So I think more work needs to be done on that, but I think we're thinking about it the right way in terms of saying what is OK to be done and what is not OK to be done. LESLIE KENDRICK MASON, JR.: So how do you deal with the fact that you've got these raw feeds that I would presume has all the information on, they're not doing little bars across the eyes while the camera's rolling. And this touches to this point about edge computing. Edge computing gives this powerful mechanism to do these controls. As long as the video is being processed at your edge, it is not leaving your premises as such. So that's a great place where the raw video can come in. You can say, for instance,

obscure or obfuscate any sort of sensitive information and then re-encode it and send it out so that whatever goes out already has a bunch of these PII's stripped out. PII's? Personally Identifiable Information. Once they are out, then what we have in the cloud is much less sensitive, if you would. And then, of course, on top of it, if you add policies in terms of what you can do and cannot do, then I think, yeah, we have designed something that's not necessarily going to enable Big Brother, but enable things that we think are constructive and useful to society. Right.

One of the works that we're doing right now, actually, is using secure enclaves for video analytics, even in the cloud, so that the only person who would know what's happening inside is the person that's actually running it. Even the cloud operator or the operating system would not be able to immediately snoop into what's happening. Well, it's story time, Ganesh. And again, you have a rather entertaining one. Tell us a bit about yourself. What got you interested in systems and networking research?

Where did you start and how did you end up here at Microsoft Research? I dabbled in a few things during my undergrad, like anybody would in terms of saying, what is it that really interested me? And, you know, it kind of got to this conclusion where at the end of my undergrad, I thought, hey, I wanted to do systems and networking. And it was, of course, in a broad sense. And so in the place I did my undergrad,

we had this requirement where the last full semester, we had to do an internship. And so the forms that you fill in, I had filled in MS R&D. It was just called Microsoft R&D in Bangalore. And I was kind of excited because I thought, hey, OK, so this is a lab that I've heard of. You know, some of my friends had interned in Microsoft Research during their undergrad. I'm like, okay, so this sounds good, and it should give me a chance to try out things. But that didn't exactly go as expected. After I landed there, I realized that this was actually a global technical support center.

This was, while it's a hugely important job. Absolutely. It was, as I said, not expected in terms of what I wanted to do. Right. But at the same time, I also figured that, you know, Dr. Anandan, who was then the head of the MSR India lab, he had just opened the lab then, and I saw his press conferences and so forth. And so what I

did was I found where his office was and I went over to his office and as the way he puts it you know I hired myself in that I essentially told him that hey we would like to try my hand at research you know can you give us a chance at it and that turned out to be pretty good. I remember at that point, he said, okay, come on in. And we started working with him and his team. That was kind of cool. And so at that point on, I, you know, I personally also decided, why don't I spend, you know, a little more time at this place after my internship ended. And so, yeah, that got me interested in systems and networking. Then I went to

Berkeley for my PhD and then here in Redmond. Tell us one thing about yourself, something interesting, a characteristic, a life event, a trade, a personal quirk, an interest of yours that we might not know about you and how it has influenced your career as a researcher. I can relate to the game of cricket. There is a form of cricket called test cricket. And believe it or not, it's for five days. So it's a really long game or a sport, if you would. And the game as such is split, you know, over five days. Each day has three sessions.

And usually the lesson I took away from cricket was that if you want to win the entire test match, that you don't formulate every session in a win-lose manner. In that, you know, at the end of each session, just make sure you're building up towards winning the match. But don't take any of the short-term things too, you know, seriously, both in a positive sense or a negative sense, because that doesn't eventually lead to you winning the match. You win the match if you keep the eye on the ball for the entire five days. So yeah, it helps me kind of, in many ways, balance between what I do short-term, what happens in the short-term, both positive and negative. But I always make sure I try to think for what it is in the long term,

or the test match, if you would. So research as test match. Research as a cricket test match. Yeah. Well, at the end of every podcast, I give my guests a chance to say anything they want to our listeners, some of whom might just be getting their feet wet in systems and networking research. So it could be something helpful, inspirational, cautionary, encouraging, profound, whatever. listeners, some of whom might just be getting their feet wet in systems and networking research.

So it could be something helpful, inspirational, cautionary, encouraging, profound, whatever. You get the last word. What would you want to say? Dr. To people that are getting their feet wet in systems and networking, I can pass on lessons that I've learned from the many illustrious people I have worked with. It's always important when you choose a problem to look at something that is relevant, ask what's the impact, whose life is it going to make better, and often you know I've seen that in cases when I have done it that things turn out to be much better than the cases when I have not done it. So that's a very handy thing to have as a question. While designing a solution,

something I always try to do is try to keep the solution simple. Do not complicate the solution. There is no great heroism in having a complicated solution. Simple is always better. And seek out collaborators. There's no better way to amplify your work. They have a huge multiplying effect when you have good collaborators. So yeah, that's what I would tell to anyone who is starting out. Occam's Razor with a great team. Yeah, that would be the way, yeah. Ganesh Ananthanarayanan, thank you for joining us on the podcast today.

Thanks for having me over. It was fun. To learn more about Dr. Ganisha Nantanarayanan and the latest in systems, networking, and mobility research, visit Microsoft.com slash research.