1.1. Paper#

Title

How to take environmental footprint into account when designing IoT solutions

1.1.1. Introduction#

IoT systems are everywhere, from finance to health, from manufacturing, to agriculture, and their number and complexity is continuing to grow[RGR18]. But while the benefits to society are great, when we put that growth in the perspective of the ever growing greenhouse gas (GHG) emissions from ICT, it is clear that we need to start to take the environmental footprint into account when designing, developing, and maintaining such systems.

An architect of IoT systems needs to take many aspects into account in his design, like latency, security, and cost, but environmental footprint is rarely one of them[GS18]. And when it is, choices are made more on the basis of belief than on facts.

Ideally the architects have tooling that aids them in their design and an IoT system framework that manages environmental footprint during use. Environmental footprint is calculated on the entire equipment life-cycle, manufacturing, distribution, use, end-of-life, but these numbers are very hard to come by, and may change significantly over time. Such a framework must encompass both design and use because the choices in design affect the capabilities during use, and the current capabilities affect the choices in the inevitable redesign.

An IoT system is not static. It is a continuously evolving system with updates, upgrades, and extensions of both software and hardware. In this article we explore how choices in software architecture affect choices in equipment, we explore the environmental footprint of equipment used in IoT systems, and how that relates to the IoT system as a whole. How to manage the environmental footprint over the entire lifetime of the IoT system and how to integrate environmental footprint into other system-level concerns is out of scope.

IOT systems are frequently used to manage equipment that consumes power. In this regard an IOT system can have a net positive effect on environmental footprint by reducing the energy consumption of the equipment it controls. This reduction is caused by the IOT application, not the IOT system that implements it. As such, these footprint reductions are out of scope.

Moreover, and where applicable, we will also limit the scope to the ecological priorities defined by the European Union.

[Contributions]

  • Deliverables zonder relatie onderling (bvb. proof of concept/prototype, echte werkomgeving, interviews, new software, …)

  • De ingredienten waarmee je gaat werken

Contributions

  • Approach / methode

    • Hoe ga je aan de contributies komen en wat zijn de afhankelijkheden. Zit er een volgorde in.

  • We will do an extensive literature study on environmental footprint of equipment in IoT

  • Life-cycle analysis of equipment

  • Large trends and statistics

  • European policy

  • …

Approach

1.1.2. Environmental footprint of ICT#

The environmental footprint of a product is the sum of the ecological impact for each stage in its life-cycle. In the EU the ecological impact is measured and weighted on 16 different categories [ZP19][Zam19], but not all categories are relevant for IT. Benqassem et al. [BBdeMontenay+21] identified 8 categories that are, with the top 3 being: climate change (greenhouse gas emissions), weighted at 21.5%; minerals and metals resource use, weighted at 6.7%; and fossils resource use, weighted at 8.3%. Greenhouse gas (GHG) emissions is deemed most important, by the EU and by others, and has received - by far - most attention ([DTU20], [BE18], [Mar20], [WeN21], and many more). The statistics and predictions are not without debate, but we can at least conclude that the GHG emissions of ICT is significant and rising, and that manufacturing of equipment plays a major, if not dominant role in these emissions.

In 2021, the European Parliament group of the Greens/EFA published a report[BBdeMontenay+21] on a more thorough study on the ecological impact of ICT in Europe. They report that manufacturing is responsible for 40.1% of the ICT GHG emissions and use 57.8%. However, they also conclude that although the impact of GHG emissions is great, the impact of resource use (minerals, metals, and fossil) in manufacturing is greater still. They advise that “[…] they should therefore be taken into account first and foremost in all strategies to reduce environmental impacts.”

[TODO: EU requirements on reporting on AI/ICT footprint.]

1.1.3. Anatomy of IOT systems#

An IOT system provides the means to efficiently and effectively manage a real-world process. Example of real-world processes include managing traffic in a city, managing crop yield in a farm, managing produce quality in a factory, and so on. To do so the system needs to sense or measure a variety of attributes at a variety of locations, deduce what is happening, make any necessary changes to the process, and report on the state of the whole. For example, in a greenhouse sensing that the inside temperature is too high, the solar radiation is good, and the outside temperature is high, the system deduces that cooling is necessary, reports the situation, and activates the equipment.

So the system uses sensors and actuators to interact with the physical world, it uses computers and storage for analysis and to make decisions, and it uses dashboards and controls to interact with people. What makes it an IOT system is that all these devices communicate over TCP/IP.

There are many forms of computing and storage that can be used in an IOT system:

  1. Cloud computing. Sensors, actuators and dashboards are connected to the internet and directly accessed and controlled from software running in a cloud data center. All data is stored in the cloud.

  2. On-premise computing. All servers are on-premise, close to the sensors, actuators, and dashboards. This is where all compute is done, including training, and where all data is stored.

  3. Edge computing. The software is partitioned into tasks that need to happen close to the sensors and actuators, and other tasks. **There are servers on-premise for the edge tasks, but other tasks are done in a cloud data center. This is also where data is stored and training is done.

  4. Fog / Mist computing. The software is partitioned as in edge computing, but rather than centralizing the edge tasks to local servers the tasks are distributed among existing local compute like sensors, actuators, and network equipment. Data is still stored in cloud data centers where training happens.

1.1.3.1. Research questions#

The question we wish to answer in this article is how to take ecological impact into account when designing IoT solutions. When does it make sense to use local edge computers, on-premise data servers, or compute services in the cloud? What are the conditions of this decision, and how frequently do they change? And, finally, would it make sense to change the location of compute dynamically, or is the rate or magnitude of change too low?

As a working scenario we use a system where the edge is comprised of several IoT sensors and actuators, and some dashboards for aggregate data; where compute is done on the data streams for monitoring, control, and aggregation, and where historical data is recorded to train new (machine learning) models.

1.1.4. Environmental footprint of hardware#

1.1.4.1. IT life cycle#

There are four life cycle phases for IT equipment, each with a unique ecological impact: manufacturing, distribution, use, and end-of-life. The ecological impact of IT is dominated by the manufacturing phase[BBdeMontenay+21], mostly because of mineral and metal use. Unsurprisingly, greenhouse gas emissions dominate the ecological impact for IT usage. While relevant in general, in our comparison we will not take into account the differences in distribution phase, useful life, and end-of-life phase. We expect the impact to be negligible, and differences are greatly influenced by varying operating conditions or organizational policies.

1.1.4.2. Manufacturing#

We must take manufacturing into account, given its impact. Throughout the lifespan of an IoT solution, IT equipment will probably be replaced several times. The main difference between the various scenarios is in where compute is done, and so in which servers are used at which location.

A major difference between on-premise servers and datacenter servers is in how peak performance demand is handled. With a data center peak performance can be managed by the elasticity of the data center. A fully on-premise server park must be scaled for the worst case performance and may have high manufacturing impact simply because it must have more servers. As such we will use the average compute load to calculate the manufacturing impact for data centers, and the worst case compute load to calculate the manufacturing impact for on-premise solutions.

You could argue that the total manufacturing impact of a datacenter dwarfs the contribution of a single IoT solution, but that does not mean it is negligible when compared to alternative solutions. We should not assume the manufacturing impact on a datacenter is zero, as this would give an unfair advantage to datacenter use.

To reduce the ecological impact of manufacturing, the solution should use as few computers as possible. It is better to increase the compute power of a computer than it is to add another low-power computer. If we were to take this to extremes we should increase the compute power at the sensors and actuators and forego additional servers. [provide numbers and references to warrant this]

1.1.4.3. Usage#

We can characterize the differences between the various solutions as follows:

  • Networking equipment utilization. The network distance between edge and compute is highest in cloud computing which will utilize the most networking equipment.

  • Idle and stand-by time. Cloud computing can be designed to allow for multi-tenant server usage and thereby reduce energy waste by reducing idle and stand-by time.

  • Power-use Effectiveness (PUE). The percentage of energy used for actual computation.

1.1.4.3.1. On-premise computing#

Studies sponsored by Microsoft, [Wal20a], and Amazon, [Biz19], and reaffirmed in [BBdeMontenay+21], indicate that cloud computing has a significantly better carbon footprint than traditional on-premise datacenters. The reasons are that traditional datacenters tend to be overprovisioned to cater for potential worst-case demands. Cloud vendors mitigate this through multi-tenancy. To quote Walsh: “Generally, as the number of users increases, the ratio of the peak demand to the average demand for the user set decreases.”

Another reason is that cloud datacenters tend to have more efficient IT equipment and datacenter infrastructure.

1.1.5. Fictional example#

Company A has an IoT solution for their traffic lights and sensors, consisting of one IoT sensor per traffic light, with the following properties:

  • Power draw of 1W (8,76 kWh per year, ~2 kg CO2 per year)

  • Production costs of 10 kg CO2

  • Expected life span of 10 years, of which 8 have elapsed.

There is a new chip available, with a power draw of 0.5W (~1 kg CO2 per year), and production costs of 20 kg CO2, with an expected lifespan of 20 years. When is it worthwhile (or optimal) to replace the chips from a sustainability standpoint?

The total expected emissions for the old sensors was 10 + 102 = 30 kg CO2. With the current lifespan that is 10 + 82 = 26 kg CO2. The new chip will costs in it expected lifetime 20 + 20*1 = 40 kg CO2, or 2 kg CO2 per year.

Stated differently: the new chip costs 1 kg less per year to operate. To earn back the initial production costs of 20 kg CO2, the new chip needs to operate for 20 years. Drawback: comparing two scenarios: one of 20 years (replace directly) and one of 22 years (wait + replace).

Conclusion: don’t replace, better to use the current hardware for the remaining 2 years? Why?

However, should we include a fixed timespan of the next 30 years and calculate for the whole timespan the optimal solution?

Replace now: 22 years 2 kg CO2 per year, assuming a similar device is bought after the lifespan to replace it. Replace later is 2 years 2 kg CO2 per year and 20 years 2 kg CO2 per year. Is the same.

However, don’t fix that ain’t broken, because it costs 20 years to earn back the production costs. So better to delay that. But how to quantify?

Is this not just a LCA? No, given a LCA, how can we incorporate this in a software design? Mainly about lifespans.

1.1.6. Notes#

  1. Edge and fog computing focuses on minimizing latency, but in real-time situations the latency often does not need to be as small as possible, it just means that there are deadlines that need to be met. What gains could we get from making edge/fog computing deadline-aware?

    1. Could SDF be used for this?

  2. https://link.medium.com/iOW6q3XrLzb

Wel of niet vervangen in scope? Waarschijnlijk niet.