Watt’s the best time to compute? Google’s road to the Edelman Award

Ana Radovanovic:

I think that the possibility to do something like this, it has been a great journey and that when you have a great idea that affects everybody in the world is just right and the right team, nothing is actually hard and it’s not getting the team that the idea gets together at people and the teamwork can really move and change the strategy of a large company like Google. And yeah, this is something that still is unbelievable probably for us that this happened and the impact that we created on the other hand, it makes us proud. And we can give this story to other data scientists, other engineers, and I think this is a great inspirational story for their own ideas.

Ashley Klimp:

Welcome to Resoundingly Human, the Informs podcast. I’m your host, Ashley Klimp, and in this special miniseries, we’re highlighting the finalists of the 2026 Franz Edelman Award, which highlights organizations from around the world that demonstrate how advanced analytics and operations research are solving real problems at massive scale. Joining me today are members of the team from Google, whose project is advancing a new approach to managing energy and computing resources across the company’s global data center network. Data centers power the digital services we rely on every day, but these also require enormous amounts of energy. To address this challenge, Google developed a first of its kind system that intelligently shifts complex computing workloads across both time and location to take advantage of cleaner energy sources. In today’s episode, we’ll hear more about how this innovative system works and the impact it’s having. So I’d love to kick things off with introductions.

If you could each introduce yourselves, including your role at Google and your involvement in this Edelman finalist project.

Ana Radovanovic:

Hi, I’m Ana Rodovanovich. I’m a technical lead for Carbon Aware computing and energy analytics at Google. And I’ve been at Google for already 18 years and mostly working on various applied research problems.

Ross Koningstein:

Hi, I’m Ross Conningstein. I was at Google for 25 years, recently retired. I was the first director of engineering and helped Ana get this project going.

Walfredo Cirne:

Hi, I’m Valfredo Sirni. I’m the TL for Flex, which runs the Google Internal Cloud where we support things like Gmail search ads and of course the external cloud runs on our internal cloud infrastructure. Flex does manage who runs where and what guarantees they have running in our internal cloud. And then it was a natural partnership to work with Anna, Adrian, Rose, and all the rest of the team to make us more efficient, not only in terms of the resources needed to build the data center, but directly the energy we consume.

Adriaan Hilbers:

And I’m Adrian. I’m the most recent member of the team here. I’ve been wanting to join the team for about five years and finally talk my way in about a year and a half ago. I’m a data scientist and I was responsible mostly for quantifying and presenting the impact that this program is having as well as being the data science lead for the demand response program where we drop data center electricity demand when requested by a grid operator.

Ashley Klimp:

Well, it’s wonderful to meet you all. Thank you so much for joining me and let’s dive into the challenge itself. Data centers operate around the clock and power services used by billions of people. What made managing energy use across such a large global infrastructure such a complex problem?

Ana Radovanovic:

Well, maybe the idea was just simple in 2018, but the system like Carbon Aware computing never existed before. Managing workloads and there are a lot of technical challenges associated with managing workloads and this is for many reasons. You have to have the infrastructure built in such a way and I would let Valfredo talk more about that, but I think just having a control and consequences of moving workloads without really negatively affecting the performance that user expect from it is not easy to do. Also, there were other challenges associated with shifting computing and this is mostly stems from the data and analytics that we need. Obviously, if we want to do something like carbon aware computing and shifting computing at to times and locations where it’s greener to do so where we can consume cleaner energy, we need to have a carbon intensity data. At the time when we started, this was actually only the start from having this data available and this is how we started really directly connected to people that could source it and there were very few places that could do that.

And in the end, accomplish something like that at scale is tremendous job. It’s a lot of work and it took us many years to get to where we are today.

Ross Koningstein:

In terms of complexity, one of the challenges is that Google’s infrastructure managing millions doesn’t even begin to touch the number of jobs that are associated. So you’re actually rearranging billions, let’s say, of workloads. So scheduling them individually is impossible. You have to mold it like Play-Doh during the day and across data centers to get things to work and meeting performance requirements is hard and that’s just the technical challenge. One of the other challenges within Google is there’s always a lot of competing things that want to be happening. And so how do we get Google to agree that we should do this project was another one of the challenges. And without going into really much detail, the first thing we had to do was prove that we could even make this happen. And so there was a small team that did that. And once we accomplished that and showed the graphs that convinced people, then we got more resources and we moved into time-based load shifting.

And when that started to work, then things grew bigger and moved to across data centers and so on. So it’s one of those projects that it took successive buy-ins to make things work and that’s just getting at the management part. Then there’s the technical fund.

Walfredo Cirne:

There is a lot of technical complexity because when you solve the problem in a more kind of academic setting, you can simplify certain things and go to the core of the problem and solve it. But when you need to deploy in practice, you cannot simplify many of these things and you just have to deal with them and then the complexity really adds up and makes things more complicated to make progress. And in the non-technical side, every scheduling decision you made is a trade-off, even besides approvals and so on, just creating the governance framework for people to reason about the trade-offs they are incurring when we also try to maximize efficiency, improve, throughput, do all the other things and reduce our carbon impact. That’s an interesting work on itself that’s kind of the other side of the coin to the technical work.

Ashley Klimp:

So next I’d love to hear, could you give us a high level overview of how Google’s carbon-aware computing system works?

Ana Radovanovic:

So having all this complexity in mind that all of us shared just a few minutes ago, we had to understand what is the most effective way to manage computing. And actually because of all these complexities and governances that Valpredo and Ross mentioned, we decided we understood that actually the power of analytics is huge in this space and that really what we need to do is to try to actually build the analytics pipelines that would understand the complexity, understand the flexibility of workloads, that we could shift in a carbon away, source the data and build the tools that are going to provide load shaping guidelines for shifting work. And this was the novelty. This was the conceptual view like shift from having only real-time systems to having a real-time systems actually controlled by something by a system suite of analytical pipelines that would really, really provide low shaping guidelines and also add these special flexibility hints to schedulers to, for example, reduce the demand, reduce the available resources in certain locations that are carbon intensive or that would increase capacity at locations where it’s greener and where we encourage workloads to go because we want to increase the consumption or the demand at locations where we need that basically have low carbon intensity.

And all of that was happening so smoothly after that in many ways, right? We have to many monitoring systems that actually track what we are doing. We have a feedback control which monitors, “Oh, did we meet all the guidelines? Did we meet all the demand that we need to meet?” And actually providing some kind of feedback for us to see whether we are doing, we are meeting all the objectives while doing this carbon aware load shaping. Obviously building all of such a system to cures. We had a lot of learnings to accumulate. Not everything worked from day one and not everything works even today. It was also like trials, a lot of trials, tests, learnings, adaptations and making the system perform

Adriaan Hilbers:

Better. We’ve had various dashboards and that’s a little bit the trick of … And you will see some of those in the presentation is how can we distill the impact that we’re having to the killer visual, so to speak? And that’s definitely a lot of work. And what we settled on, and we’re going to show this in the presentation as well, but what we found a very useful tool is to look at individual data centers and say, how much did we change the total amount of compute load going into that data center? So you might say, “Okay, we ran the equivalent of 100 CPU hours or 1000 CPRs less in this data center and 1000 more somewhere else.” And basically the sum of those deltas is always zero, right? Because we’re only moving load around and we’re not sort of creating or destroying it to use the physics analogy.

But if you make that same plot, which is not the delta in compute usage, but the delta in carbon footprint, actually there’s a bias, right? We tend to move compute load to places where the carbon efficiency is more efficient so the carbon intensity is lower. So actually while the sum of all the changes in compute load add up to zero, the sum of all the changes in carbon footprint add up to something negative, which is that we’ve reduced the carbon footprint. I think that sort of visual that we settled on is the most clear intuition of our impact.

Ana Radovanovic:

Yes. I think thanks to Adrian, finally when he joined our team, we finally had an amazing way to show our impact and to convince leadership that we are saving really what we are saying. We were always having one shot graphs showing this, but now we have real time monitoring of day-to-day savings in energy consumed and carbon footprint. Yes.

Ross Koningstein:

I think one of the benefits here that maybe has been glossed over a bit is that this is a day ahead predictive system. That means that it can make decisions that it can see ahead of time. And a good example of this is peak capacity needs. So if you just sort of randomly throw load at computers, what we call organic load, you can have peaks during particular times of the day, but those peaks, part of that load is actually not important and it can be moved somewhere else. And when you do that, you can reduce the number of data centers you actually need to have or need to build as the case may be. So if you factor in that you’re trying to also peak shave, as we call it, then that helps the company save money as well as reducing carbon load.

Ashley Klimp:

So one particularly interesting aspect of the project, and I think we’ve touched on this a litle bit already, is the ability to shift flexible compute workloads across both time and geographic location. How does the system determine when it makes sense to move workloads in order to take advantage of cleaner energy sources?

Ana Radovanovic:

This actually happens really using this proactive, meaning the ahead system. So we build analytics that understand what’s going to happen on the grid, that understand the next day demand, that understand which workloads can be shifted and which not. And so thanks to this, we provide shaping guidelines and the system knows really because it’s not a random … Obviously there are random effects, but a lot of things we learn and we always try when we learn to reduce the uncertainty in making decisions of that kind. So we keep the control of our system under robust. And so this is how we know. So we source data every day, we run and learn and do forecasting. We do this day ahead planning. We produce signals and this way we control systems. Anything obviously, anything that happens that is unpredictable, we actually manage to catch on time. So obviously unpredictable things can happen and we cannot avoid them, but we are trying to avoid them in a way that we provision our systems and controls to catch them at the right time and respond to that.

Walfredo Cirne:

I think this particular system has a very interesting combination of a datahead lookout where we can benefit from the statistical behavior of the workload. We run many, many million jobs a day in our data centers. And although every single job is hard to predict exactly what they do, when we aggregate, then we have a much easier time. So we’ve run a very heavy analytics that predicts the next day, but that guides the online scheduler because you can imagine the online scheduler that is deciding where I place a job now, do I delay this job for an hour or two such that we can run this job under more greener energy and many other scheduling criteria that goes along that. The scheduler is dealing with an incredible number of jobs and the scheduler just cannot sit thinking for 15 minutes what to do for that particular job.

First of all, it already automatically delayed the job even before you making a decision, but also you need to process the millions and millions of jobs. So we split the problem such that in the previous day we actually compute the shape or what’s the aggregated load every single minute should look like to inform the online scheduler that is doing the job that’s making decisions very fast, where a job should go, how much they should wait if anything at all to be informed by the shape of the load that we want to create.That’s a very interesting technical solution to be able to deploy. We just do not have the time to be running very complex computation that would take minutes on every job

Ashley Klimp:

All right, let’s talk about the impact. Since 2020, this system has been active across Google’s global data centers and has helped position the company as a leader in carbon-aware computing. What impact has this had on Google’s sustainability efforts and overall data center operations?

Ana Radovanovic:

Yes. So the impact was gradual and I would call this as a movement. This didn’t exist. As an idea, the system itself and this type of an effort at such a large scale with gradual deployment of first time shifting and then eventually spatial shifting and actually learning about other benefits, not only for to align computing with less carbon intensive resources, cleaner resources, generation on the grid, but also to move workloads to more power efficient clusters and reduce the energy we consume, then to use the same type of mechanism to shadow when the grid is under stress and to help 20,000 homes keep their electricity during the storm in Oklahoma in 2021 so the applications and actually use cases grew and people were coming, this was really nice approaching us seeing the effect that this load shifting could have on so many objectives. So starting with just carbon and power, it was just really like peak load shaving and power shedding and then it was just really efficiency, right shifting workloads to reduce trending, to better use our resources, to more effectively use the network links because we choose to shift workloads that are cheap, consume less network bandwidth.

So I think this is one side of the story. And then what we decided also to stay committed to also our sustainability objective and to see how we can share this with the world and what is unique about our project that we wanted, we wanted to kind of spread knowledge to make it implementable elsewhere, to share the more information about which signal grid signal we need to follow to provably reduce carbon dioxide emissions from the dispatch generation on the grid. And there is a research on that. There is research on power models that we created to map computing demand to power. There is a research that describes the system that we build in hope that others compute providers could do something similar or come up with their own way to do something similar. And I think this is really growing even today the application of this domain.

I think during the presentation we have people who can witness and say more how we impacted the world from the outside world. But what happened is also that we haven’t stopped. This is not done. We are now scaling the project to shift more load segments, but also we are trying to do research and external collaboration to see what flexibility means for electricity grid. And in these times where electricity grids are under huge stress and load and then infrastructures you cannot build overnight, we are trying to expand what we do and really understand this impact of load shaping and shifting on more sustainable, affordable and efficient grid operations.

Ross Koningstein:

So I’d like to add one more dimension in terms of impact and that has to do with just how happy people are to work on the project. So when we started this project, it was like there was a group of people within Google sustainability organization that’s like, “We want to be working on this. ” And that kind of positive, I don’t know, esprit decor really brings out a lot of creativity in people and it’s great. And once we started working with our external partners and I’m talking about the analytic data provision for the CO2 information grid modeling, the different point of view that we were all taking and looking at really got people excited about the possibilities. And so I think this effort has benefited by a lot of creative input from a lot of people because they were genuinely excited about what we could do.

Ashley Klimp:

The Franz Edelman Award celebrates analytics that make a real world impact. When people hear the story of this project, what do you hope they take away about the role of operations, research and analytics in building more sustainable technology infrastructure?

Ana Radovanovic:

I think that the possibility to do something like this, it has been a great journey and that when you have a great idea that affects everybody in the world and the right team, nothing is actually hard and it’s not getting the team, that the idea gets together at people and the teamwork can really move and change the strategy of a large company like Google. And yeah, this is something that still is unbelievable probably for us that this happened and the impact that we created. On the other hand, it makes us proud and we can give this story to other data scientists, other engineers and I think this is a great inspirational story for their own ideas.

Ashley Klimp:

I want to offer my congratulations to you and the rest of your team from Google. I’m being named a Franz Edelman Award finalist. Thank you so much for joining me today and I look forward to meeting you in person at the upcoming 2026 Informs Analytics Plus conference where all the finalist teams will present their work during the competition and celebrate the winner of this and other Informs Awards of the Edelman Gala. To learn more about this project and other finalist teams, visit resoundinglyhuman.com and to all our listeners and viewers, be sure to like and subscribe wherever you watch or listen, the Resoundingly Human podcast.