Ideas on a lower-carbon internet through scheduled downloads and Quality of Service requests
Other titles:
- The impact of internet use and what we might do about it?
- Opportunities for powering more internet use with renewables
- I want this thing, but not until later
- A story of demand-side prioritization, scheduling and negotiation to take advantage of a fluxuating energy supply.
I recently got interested in how renewable power generation plays into the carbon footprint of internet usage. We need power to run and charge the devices we use to consume internet content, to run the networks that deliver that content to us, and to power the servers and data centers that house those servers.
Powering the internet eats up energy. The power necessary to serve up the files, do the computation, encode and package it all up to send it down the wire to each of the billions of devices making those requests consumes energy on an enormous scale. The process of hosting and delivering content is so power hungry, the industry is driven to large extent by the cost and availability of electricity. Data centers are even described in terms of the power they consume - as a reasonable proxy for the capacity they can supply.
One of the problems we hear about constantly is that the intermittent and relatively unpredicatable nature of wind and solar energy means it can only ever make up a portion of a region’s electricity generation capacity. There’s an expectation of always-on power availability; regardles of the weather or time of day, a factory must run, a building must be lit, and if a device requests some internet resource the request must be met immediately. So, we need reliable base load generation to meet most energy demands. Today, that means coal, natural gas, nuclear and hydro generation plants - which can be depended on to supply energy day and night, all year round. Nuclear and hydro are low-carbon, but they can also be expensive and problematic to develop. Wind and solar are much less so, but as long as their output is intermittent they can only form part of the solution for de-carbonizing electricity grids across the world - as long as demand not supply is king.
There are lots of approaches to tackling this. Better storage options (PDF) smooth out the intermittency of wind and solar - day to day if not seasonally. Carbon capture and sequestration lower the carbon footprint of fossil fuel power generation - but raise the cost. What if that on-demand, constant availability of those data centers’ capacity was itself a variable? Suppose the client device issuing the request had a way to indicate priority and expected delivery time, would that change the dynamic?
Wind power tends to peak early in the morning, solar in the afternoon. Internet traffic is at its highest during the day and evening, and some - most - is necessarily real-time. But if I’m watching a series on Netflix, the next episode could be downloaded at anytime, as long as its available by the next evening when I sit down to watch it. And for computational tasks - like compiling some code, running an automated test suite, or encoding video - sometimes you need it as soon as possible, other times its less urgent. Communicating priority and scheduling requirements (a.k.a Quality of Service) from the client through to the infrastructure used to fullfill a request would allow smarter balancing of demand and resources. It would open up the door to better use of less constant (non-baseload) energy sources. The server could defer on some tasks when power is least available or most expensive, and process them later when for example the sun comes up, or the wind blows. Smoothing out spikes in demand would also reduce the need for so-called “peaker” plants - typically natural gas power plants that are spun up to meet excess energy demand.
“Kestler: While intermittent power is a challenge for data center operations, the development of sensors, software tools and network capabilities will be at the forefront of advancing the deployment of renewables across the globe. The modernization of the grid will be dependent on large power consumers being capable of operating in a less stable flow of electrons.
What’s Ahead for Data Centers in 2021
Google already experimented with some of this, and its a fascinating and encouraging read.
“Results from our pilot suggest that by shifting compute jobs we can increase the amount of lower-carbon energy we consume”
Our data centers now work harder when the sun shines and wind blows
There are clearly going to be hurdles for wide-scale adoption of this kind of strategy, and its never going to work for all cases. But with a problem at this scale, a solution that shaves off 1%, or a fraction of 1% can still translate into huge monetary and carbon savings. So, what would it take? Are there practical steps that us non-data-center-operators can take to facilitate this kind of negotiation betweeen the client and the massive and indifferent upstream infrastructure that supports it?
The low hanging fruit in this scenario is video streaming. It represents an outsized percentage of all internet traffic - and data center load. Netflix alone generates 15% of all global internet traffic. What if even 1% of that could be shifted to be powered entirely by renewable energy, by virtue of the deferred-processing at the supply-side, or scheduled download at the client-side? Often its the case that when I click to watch video, I need it right there and then - perhaps it is a live event, or I didn’t know I needed it until that minute. Sometimes not though. If it was possible to schedule the download ensuring it was there on my device when I did need it, the benefits would ripple through the whole system - content delivery providers would save money and maybe the grid itself would be able to absorb more intermittent renewable generation.
There are other opportunities and I don’t want to get too hung up on specifics. But the notion of attaching Quality of Service in some way to some requests to facilitate smarter utilization of seasonal, regional and weather-dependent energy generation fluxuations seems promising to me. Fundamentally, power demand from worldwide internet traffic is extremely dynamic. We can better meet that demand with equally dynamic low and zero carbon sources if we can introduce patterns and signals at all levels of the system to allow it to plan and adapt.
…
When I get to the end of a piece like this I’m always left wondering “what is the point?”. Is this just a rant into the void, hoping someone listens? Its certainly not an actionable plan for change. Writing it down helps me process some of these ideas, and I hope it starts conversations and prompts you to spot these kind of indirect opportunities to tackle climate change. And if you are in a position to nudge any of this towards really existing in the world, that would be great. I work at Mozilla, we make a web browser and have our own substantial data-center and compute-time bill. I’ll be looking into what change I can help create there.
Some References
I collected a large list of papers and articles as I looked into this. Here’s a smaller list:
- Netflix Environmental Social Governance Report (2019) (PDF)
- “In 2019, Netflix’s direct energy use was about 94,000 megawatt hours” (Direct energy usage, not including cloud services)
- Content delivery network includes “We partner with over a thousand ISPs to localize substantial amounts of traffic” so its partly local
- “indirect energy use was about 357,000 megawatt hours in 2019”
- 167 million subscribers in 2019. Hours downloaded? (we have that 2011 study which claims 3.2billion hrs)
- Estimating a Data Center’s Electrical Carbon Footprint
- “Avoided emissions reflect the average activity of peaker plants in the local utility’s network”
- Sandvine: Global Internet Phenomena Report 2018
- Video is 58% of internet downstream traffic volume.
- Netflix is 15% of all internet downstream traffic
- Factcheck: What is the carbon footprint of streaming video on Netflix?
- Carbon-aware Load Balancing for Geo-distributedCloud Services (PDF)