Share
Patterns of sustainability - Going green in IT
This slide deck addresses four questions:
1. Is Green IT a relevant topic? (Yes, it is)
2. Is Software a significant contributor to the ecological footprint of IT? (Yes, it is)
3. What can we do as software engineers working in different roles to reduce the ecological footprint of IT? (A lot)
4. Is Green IT the only relevant topic of the upcoming years? (No, it is part of a bigger challenge)
As a lot of studies and presentations already exist that address the first and second question in depth, this slide deck keeps these parts short.
The main focus is on the third question. Here, a lot of patterns are presented that software engineers can use in different roles to reduce the ecological footprint of IT.
Finally, the topic "Green IT" is embedded in a bigger picture describing the core challenges, IT faces in a more and more uncertain VUCA world while becoming more and more indispensable at the same time. Still, as the focus of this presentation is on the patterns of the third part, this part is also kept short.
At the end, a lot of references that were used when preparing this presentations are included for everyone who would like to dig deeper into the topic.
As always, the voice track is missing. Hopefully, this slide deck still gives you some useful ideas to ponder ...
Uwe Friedrichsen
Transcript
159 slides
-
Patterns of sustainability Going green in IT Uwe Friedrichsen –
codecentric AG – 2020-2023 -
“The IT sector accounts for 4% of global GHG emissions
and this could double by 2025” Source: Center for sustainable systems, Green IT Factsheet (https://css.umich.edu/publications/factsheets/built-environment/green-it-factsheet) -
“Software can also be an enabler of climate solutions. Software
can be built to help accelerate decarbonization across all sectors in industry and society.” Source: Green Software Foundation, What is green software? (https://greensoftware.foundation/articles/what-is-green-software) -
But Green IT is a problem of hardware manufacturers and
datacenter designers and operators only, isn't it? -
Patterns of sustainability (Software engineering perspective) Hardware manufacturing & data
center design * Somebody Else’s Problem (https://en.wikipedia.org/wiki/Somebody_else%27s_problem) S.E.P. * We are done! 🥳 -
“The hope is that the progress in hardware will cure
all software ills. However, a critical observer may observe that software manages to outgrow hardware in size and sluggishness.” (Wirth’s law) Source: Wikipedia, Wirth’s law (https://en.wikipedia.org/wiki/Wirth%27s_law) -
“Software efficiency halves every 18 months, compensating Moore's Law” (May's
law) Source: Wikipedia, Algorithmic efficiency (https://en.wikipedia.org/wiki/Algorithmic_efficiency) -
Patterns of sustainability (Software engineering perspective) Hardware manufacturing & data
center design * Somebody Else’s Problem (https://en.wikipedia.org/wiki/Somebody_else%27s_problem) S.E.P. * We are done! 🥳 -
Requirements (Responsible: Product manager, architecture) Offline first Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Balance quality of service (QoS) • Counterforce to other quality
goals, e.g. • Availability • Response time • Quality of response • Goal is to find a sensible balance between • QoS quality goals • Required resource/energy consumption • Supported by design latency-tolerant user interfaces -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Fight feature creep • Feature creep is wasteful in many
ways • including resource and energy consumption • Validate feature viability before implementation • Monitor feature usage and remove rarely used features • Give users option to deactivate unneeded features • Also supports maintainability, evolvability, robustness, ... -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Mind footprint of requirements • Make estimated ecological footprint of
requirements explicit • Requires good measuring of energy consumption • Requires approach to estimate footprint of new requirements • Umbrella pattern for other requirements-related patterns • Also consider footprint of “nice-to-have” requirements • E.g., UI animations – looking nice, but require lots of energy • Supported by make energy-consumption visible -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Optimize for temporal decoupling • Temporally decouple request, processing and
response • Functional design issue • Technology only augments it • Supports maximize hardware utilization • Supports delay execution of compute-intensive tasks to times of low carbon intensity -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Data austerity • The more data you store and process,
the more • storage capacity you need • memory you need • compute power you need • network bandwidth you need • Store and process only the data you really need • Reduces environmental footprint significantly • Side effect: Improves security and GDPR compliance • Counterforce to widespread “big data” handling practice -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Offline first • Targeted at mobile applications • Do not
require a continuous connection to the backend • Leads to different designs • Less network traffic and bandwidth required • Often also less data required overall • Supports data austerity • Supported by optimize for temporal decoupling -
Offline first Requirements (Responsible: Product manager, architecture) Mind the energy-consumption
of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Offline first Mind the energy-consumption of functional requirements Optimize for
temporal decoupling Fight feature creep Data austerity Balance Quality of Service Requirements Patterns of sustainability (Software engineering perspective) Hardware manufacturing & data center design Design -
Design (Responsible: Architecture, sometimes in coordination with product manager) Design
carbon-aware applications Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Minimize redundancy • Redundancy often used as default to improve
availability • Multiplies resource consumption • Leads to poor utilization (unless combined with autoscaling) • Only use redundancy if really needed • Evaluate how high availability really needs to be • Ponder alternative approaches (e.g., Erlang’s “let it crash”) • Supported by balance quality of service (QoS) • Supported by minimize startup time -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Remove unneeded components • Continuously monitor usage patterns of system
landscape • Remove components that are not used anymore • Also consider removing rarely used components • Results in less overall resource consumption • Also leads to less technical debt and improved maintainability, evolvability, robustness, and more -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Design carbon-aware applications • Build carbon-awareness into applications • Make
aware of own current energy consumption • Connect with current carbon-intensity of power supplier • Use to build in optimization of own ecological footprint, e.g. • Reduce QoS if ecological footprint becomes too high • Shift workloads to times of lower carbon intensity • Shift workloads to data centers with lower carbon intensity • Supported by balance quality of service (QoS) • Supported by optimize for temporal decoupling -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Design latency tolerant UIs • Design user interfaces to support
progressive data loading • Pre-render structure, fill in data later • Load visible parts first, load other data later • Design UIs that compute-intensive data can be loaded later • Reduces pressure for very short response times • Supports reducing ecological footprint • Supported by balance quality of service (QoS) • Supported by optimize for temporal decoupling -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Slice services by resource usage • Split functionality by resource
consumption • Turn resource-intensive functionality in dedicated services • Group resource-economical features apart in other services • Frequency of use can be used as sub-criterion for slicing • Allows for better utilization of underlying resources • Additional force in functional service design • Supports maximize utilization -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Avoid polling • Polling continuously needs energy even if no
new data exists • Wastes energy • Impedes implementation of patterns like scale to zero • Use push mechanisms instead wherever possible • Supports scale to zero -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Reduce transmitted data • Sending data over the network costs
a lot of energy • Only send the data that is really needed • Do not send unneeded data because it is easier • Do not send data that “might be needed in the future” • Data not sent is better than the best data compression • Avoid chatty communication protocols and frameworks • Counterforce to other design forces • Supported by compress data at transit and rest -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Delete unneeded data • Delete data you do not need
anymore • Requires less storage, memory, compute and network • Reduces environmental footprint • Side effect: Improves security and GDPR compliance • Implement data lifecycle and retention policies • Makes clear which data to keep and which data to delete • Supports data austerity -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Keep data local • Loading data over the network costs
a lot of energy • Keep the data you need local • Do not load it from remote sources • Temporally decouple data synchronization needs • Requires carefully considered data architecture • Makes design and implementation more complicated • Supports reduce transmitted data -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Use optimized data stores/formats • Data stores and formats optimized
for the use case • reduce storage and memory needed • allow for optimized data access patterns, reducing compute and network needed • Consider, e.g., columnar data stores and/or Apache Parquet • Always carefully consider the use case • Offset potential data duplication footprint -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Cache static data • Repeatedly loading the same data costs
a lot of energy • Cache static data locally instead of continuously reloading it • Trades memory consumption for network traffic • Requires certain access frequency to become energy-positive • Cached data may become outdated • Cache coherency challenges need to be handled • Supports keep data local -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Prefer lazy loading • Eagerly loaded data often is not
used • Only load data when it is actually needed • Only load web page content when it is displayed • Only load table data when it is displayed • ... • Supports reduce transmitted data • Supported by balance quality of service (QoS) -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
“Hardware manufacture and use dominate the carbon output of personal-computing
companies (e.g., Apple). More emissions come from designing and manufacturing integrated circuits (e.g., SoCs, DRAM, and storage) than from hardware use and mobile energy consumption.” Source: Udit Gupta et al., "Chasing Carbon: The Elusive Environmental Footprint of Computing", https://arxiv.org/abs/2011.02839 -
Mind the impact on user devices • Design applications that
also run on older user devices • Test operability on older devices • Measure mobile efficiency * • Use managed device farms for testing • Test with older hardware/browsers and alike • Reduces environmental footprint significantly • Often ignored for various reasons * see, e.g., https://mobile-efficiency-index.com/en/ -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Programming languages differ widely with respect to energy, time and
memory consumption -
Source: Pereira et al. "Energy Efficiency across Programming Languages: how
do energy, time, and memory relate?”, https://dl.acm.org/doi/10.1145/3136014.3136031 -
Choose the right language • Programming languages differ regarding •
Energy consumption • Time consumption • Memory consumption • ... • Consider using energy-efficient programming languages • At least for the most resource-intensive parts of workload • Creates additional force regarding language choice -
Design carbon-aware applications Design (Responsible: Architecture, sometimes in coordination with
product manager) Delete unneeded data Choose the right programming language Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Cache static data Mind the impact on user devices Avoid polling Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces -
Design carbon-aware applications Delete unneeded data Mind the impact on
user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Design Requirements Patterns of sustainability (Software engineering perspective) Implementation Hardware manufacturing & data center design Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
“Software efficiency halves every 18 months, compensating Moore's Law” (May's
law) Source: Wikipedia, Algorithmic efficiency (https://en.wikipedia.org/wiki/Algorithmic_efficiency) -
“In ubiquitous systems, halving the instructions executed can double the
battery life and big data sets bring big opportunities for better software and algorithms: Reducing the number of operations from N×N to N×log(N) has a dramatic effect when N is large ... for N = 30 billion, this change is as good as 50 years of technology improvements.” Source: David May, The return of innovation (https://web.archive.org/web/20160303181322/http://www.cs.bris.ac.uk/~dave/iee.pdf) -
Algorithms and data structures • Use the best algorithms and
data structures for the task • Avoid O(n2) or worse, aim for O(n log(n)) or better • Sometimes intermediate transformation steps can help • Balance random access and memory consumption • Can reduce environmental footprint significantly • Foundations more and more often not known • Effects of library calls often not understood Erik D. Demaine et al., "Energy-Efficient Algorithms" (https://arxiv.org/abs/1605.08448) -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Economize on libs & frameworks • Oversized libraries and frameworks
waste resources • And often also increase cognitive load of developers • And increase the attack surface • Use as few libraries and frameworks as possible • Prefer lightweight libraries and frameworks • Implement simple solutions on your own • Counterforce to widespread development best practices -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Optimize wasteful code • Continuously monitor resource usage of application
parts • Identify parts with very high resource consumption • Optimize these parts, e.g., • Find better algorithms and data structures • Avoid unneeded sorting, formatting, transformations, ... • Consider hardware acceleration for extremely wasteful parts • Note: Avoid falling in the premature optimization trap -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Mind the Minifesto * • The most energy-efficient code is
code not written • Avoid everything not needed to solve the task, e.g. • Do not over-design • Do not implement ahead • Strive for simple solutions • Simple does not mean easy • Simple usually is a lot harder to achieve than complicated * http://minifesto.org/ -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Minimize startup time • Long startup times are wasteful •
Impedes fast, demand-driven application scaling • Results in overprovisioning of capacity • Strive for short startup times • Fosters resource-economical designs • Supports auto-scale resources • Supports scale to zero -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Scan for vulnerabilities • Attackers often want access to foreign
resources, e.g., for • Crypto-mining • DDoS attacks • Wastes a lot of energy for malicious purposes • Minimize risk by regular scanning for vulnerabilities • Best make it part of your CI/CD pipeline • Do not forget the other parts of a good security practice -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Shift left carbon-efficiency • Pondering carbon-efficiency after deployment is too
late • Hard to change ecologically questionable decision during • Requirements gathering • Software design • Software development • Software build and deployment • Make carbon-efficiency part of your quality goals • Take goal satisfaction into account from the very beginning -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Do not build on each code change • CI/CD pipelines
require energy • The more comprehensive, the more energy they consume • The more often run, the more energy they consume • Running the CI/CD pipelines less often saves energy • Maybe do not run the pipeline for each minor change • Relevant regarding “commit early, commit often” practices • Counterforce to widespread development best practices -
Implementation (Responsible: Development, architecture) Mind the Minifesto Make carbon-efficiency part
of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks -
Mind the Minifesto Make carbon-efficiency part of your software development
and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks Implementation Design Requirements Patterns of sustainability (Software engineering perspective) Runtime Hardware manufacturing & data center design Design carbon-aware applications Delete unneeded data Mind the impact on user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Source: Barroso et al., The Case for Energy-Proportional Computing, IEEE
Computer Dec 2007, pp. 33-37, vol. 40 -
Maximize hardware utilization • Always watch server utilization • Run
workload on fewer servers with higher utilization • Avoid servers with low utilization • Also ponder development environments, CI/CD pipelines, ... • Supported by slice services by resource usage • Supported by use managed services • Supports buy only the hardware you need -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Monitor power consumption • Always monitor power consumption, using, e.g.
• Public cloud: https://www.cloudcarbonfootprint.org/ • Kubernetes: https://sustainable-computing.io/ • and more: https://github.com/Green-Software-Foundation/ awesome-green-software • Supports make energy-consumption visible • Supports delay execution of compute-intensive tasks to times of low carbon intensity -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Auto-scale resources • Auto-scale resources up and down based on
demand • Usually works best in public cloud environment due to optimized resource elasticity • Can be applied to all types of resources • E.g., Scale data storage on demand (do not overprovision) • Different resource types may need different scaling strategies -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Scale to zero • Extension of auto-scale resources • “The
most energy-efficient system is the one not running” • Consider switching off application services if not used • Only run a gateway that launches services if needed • Supported by minimize startup time • Must be supported by data center operator for best efficiency • Usually supported by “serverless” out of the box -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Base scaling on significant metrics • Scaling is often based
on default metrics (CPU, memory, ...) • Not always suitable to reflect actual scaling needs • Use significant metrics to control scaling, e.g. • Request processing time (also applicable to events) • Number of queued requests/events • Number of active concurrent users • Might require exposing custom metrics to auto-scalers • Supports auto-scale resources -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Compress data at transit and rest • Data requires resources
at transit and rest • Compress data to reduce resource consumption • Includes minimizing web assets • Reduces storage and network footprint • Do not forget to offset (de)compression footprint -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Shed low-priority traffic • High-traffic situations lead to high resource
consumption • Shed low-priority requests to reduce resource demands • Not applicable to all use cases • Can help to reduce ecological footprint • Especially useful in times of high carbon intensity • Complements auto-scale resources -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Scan for vulnerabilities • Attackers often want access to foreign
resources, e.g., for • Crypto-mining • DDoS attacks • Wastes a lot of energy for malicious purposes • Minimize risk by regular scanning for vulnerabilities • Best make it part of your CI/CD pipeline • Do not forget the other parts of a good security practice -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Deduplicate data • Duplicated data requires additional resources • Deduplicate
data to reduce environmental footprint • Reduces storage needs • Requires less storage hardware to be built • Do not forget to offset deduplication efforts footprint -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Delay carbon-intensive tasks • Carbon-intensity of electricity production varies based
on location and time • Shift compute-intensive tasks to times and locations with lower carbon-intensity • Use, e.g., API of https://www.electricitymaps.com/ for tracking carbon-intensity of electricity production • Supported by optimize for temporal decoupling • Supported by design carbon-aware applications -
Runtime (Responsible: Operations, sometimes in coordination with architecture) Base scaling
on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero -
Base scaling on significant metrics Delay execution of compute-intensive tasks
to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero Runtime Implementation Design Requirements Patterns of sustainability (Software engineering perspective) On-Premises Hardware manufacturing & data center design Mind the Minifesto Make carbon-efficiency part of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks Design carbon-aware applications Delete unneeded data Mind the impact on user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware,
infrastructure and applications Buy only the hardware you need -
On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware,
infrastructure and applications Buy only the hardware you need -
Shutdown unneeded parts • Unused hardware, infrastructure, applications waste energy
• Shut down all parts not used • Turn off hardware reserved for peak load or future demand • Includes, e.g., turning off test environments while not needed • In cloud environments, shut down all instances not used • Does not only reduce costs, but also your carbon footprint • Complements maximize hardware utilization -
On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware,
infrastructure and applications Buy only the hardware you need -
“For data-center operators and cloud providers, most emissions are capex-related
– for example, construction, infrastructure, and hardware manufacturing.” Source: Udit Gupta et al., "Chasing Carbon: The Elusive Environmental Footprint of Computing", https://arxiv.org/abs/2011.02839 -
Buy only the hardware you need • Often, more hardware
is bought than currently needed • To satisfy peak loads • To be prepared for unexpected future demands • Producing the hardware has a big environmental footprint • Usually exceeds environmental footprint at runtime • Do not buy hardware currently not needed • Especially relevant for on-premises data centers • Cloud providers can better balance diverse demand patterns -
On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware,
infrastructure and applications Buy only the hardware you need -
Buy energy-efficient hardware • There are big differences in energy-efficiency
of hardware • Buy hardware with best W/Tx ratio for your use cases • Can significantly reduce your carbon footprint • Complements maximize hardware utilization • Complements buy only the hardware you need -
On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware,
infrastructure and applications Buy only the hardware you need -
Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications
Buy only the hardware you need On-Premises Runtime Implementation Design Requirements Patterns of sustainability (Software engineering perspective) Public cloud Hardware manufacturing & data center design Base scaling on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero Mind the Minifesto Make carbon-efficiency part of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks Design carbon-aware applications Delete unneeded data Mind the impact on user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Move workloads to public cloud • Public cloud providers put
much effort in energy-efficiency • More than on-premises data centers can afford • Can achieve much better utilization of hardware • Provide many options for reducing energy consumption • Support observing energy consumption • Yet no altruism, but part of their a business model • But you get better support than possible in on-premises DCs -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Use managed services • Complements move workloads to the public
cloud • Managed services typically allow for optimizing utilization • Also supports scale to zero out of the box • Supports maximize hardware utilization • Supports scale to zero -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Follow the carbon intensity • Data centers have different carbon
intensity • Some data centers use more clean energy by default • Some data centers have varying patterns over the day • Cloud providers usually provide data about carbon intensity • Shift workloads to data centers with lower carbon intensity • Balance with reduce transmitted data • Balance with keep data local -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Use adequate storage classes • Different storage classes use different
amounts of energy • SSD require more energy than HDD • HDD require more energy than archival-class storage • ... • Align data access frequency and storage classes used • Complements data austerity -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Move close to the users • Reducing network traffic reduces
your ecological footprint • Move closer to the users, e.g. • Choose a region that is closer to the users • Use a CDN or leverage edge computing • Can be applied to data and to workloads • Supports reduce transmitted data • Balance with keep data local -
Public cloud (Responsible: Architecture, operations) Shift workloads to data centers
with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users -
Shift workloads to data centers with lower carbon intensity Move
workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users On-Premises Runtime Implementation Design Requirements Public cloud Patterns of sustainability (Software engineering perspective) Cross-cutting Hardware manufacturing & data center design Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need Base scaling on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero Mind the Minifesto Make carbon-efficiency part of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks Design carbon-aware applications Delete unneeded data Mind the impact on user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Make energy-consumption visible • Make energy-consumption visible at all levels
• Make visible to business experts (e.g., at feature level) • Make visible to software engineers (e.g., for apps and CI/CD) • Make visible to users (e.g., at feature level) • Transparency supports balancing feedback loops • Visibility leads to changed behavior and usage patterns • Visibility leads to different decisions • Does not require explicit “education” or alike -
Continuously improve • Becoming ecological sustainable is not a one-time
effort • It is a long journey without an end state • You will not be able to do everything at once • Constantly reserve time and resources • Make sustainability a continuous improvement process -
Do not feed the resource troll • Avoid resource-hungry solutions
• Try not to use them • Try not to recommend them • E.g., be reserved regarding resource-hungry AI solutions * • Extensive use fosters disproportional energy consumption • Only use if really needed • Prefer (AI) solutions with comparatively smaller footprint * see, e.g., “Deep learning’s diminishing returns”, https://spectrum.ieee.org/deep-learning-computational-cost or “The Carbon Footprint of ChatGPT”, https://towardsdatascience.com/the-carbon-footprint-of-chatgpt-66932314627d -
Continuously improve Make energy- consumption visible Do not feed the
resource troll On-Premises Runtime Implementation Design Requirements Public cloud Cross-cutting Patterns of sustainability (Software engineering perspective) Hardware manufacturing & data center design Shift workloads to data centers with lower carbon intensity Move workloads to the public cloud Align data access frequency and storage classes Use managed services Move close to the users Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need Base scaling on significant metrics Delay execution of compute-intensive tasks to times of low carbon intensity Maximize hardware utilization Monitor power consumption Shed low-priority traffic Compress data at transit and rest Deduplicate data Auto-scale resources Scan for vulnerabilities Scale to Zero Mind the Minifesto Make carbon-efficiency part of your software development and test process Use efficient algorithms and data structures Optimize wasteful code Do not run builds after each minor code change Minimize startup time Scan for vulnerabilities Economize on libraries and frameworks Design carbon-aware applications Delete unneeded data Mind the impact on user devices Remove unneeded components Keep data local Minimize redundancy Use data stores and formats optimized for the use case Avoid polling Choose the right programming language Cache static data Prefer lazy loading Reduce transmitted data Slice services by resource usage Design latency-tolerant user interfaces Offline first Mind the energy-consumption of functional requirements Optimize for temporal decoupling Fight feature creep Data austerity Balance Quality of Service -
Core messages • Do less • Optimize at all levels
• Ensure observability • Counterforce • Nice side effect: Smaller attack surfaces * * https://portswigger.net/daily-swig/lean-green-coding-machine-how-sustainable-computing-drive-can-reduce-attack-surfaces -
Dimensions of sustainability • Ecological sustainability – going green in
IT • Economical sustainability – surviving in the long run • Technical sustainability – creating dependable systems • Social sustainability – sustainable for everyone involved -
Core IT challenges of the 21st century (Responsible: Everyone) Resilience
Sustainability Simplification -
Summing up • Ecological sustainability has become an important IT
topic • IT is responsible for more and more GHG emissions • IT can support decarbonization in other sectors • Software is a big driver of IT GHG emissions • Software engineers can positively influence GHG emissions • Requires improvements from requirements to production • Many patterns available • Often counterforce to widespread “best practices” • Ecological sustainability is part of a bigger challenge -
A journey of a thousand miles begins with a single
step -- Chinese proverb -
References and further reading Foundations & core principles • Principles
of Green Software Engineering https://principles.green/ • Greenhouse Gas Protocol https://ghgprotocol.org/ • Green Software Foundation https://greensoftware.foundation/ • Software Carbon Intensity (SCI) specification: https://github.com/Green-Software-Foundation/ software_carbon_intensity/blob/main/Software_Carbon_Intensity/ Software_Carbon_Intensity_Specification.md -
References and further reading Foundations & core principles • Gupta
et al., “Chasing Carbon: The Elusive Environmental Footprint of Computing” https://arxiv.org/abs/2011.02839 • Barroso et al., “The Case for Energy-Proportional Computing” https://www.researchgate.net/publication/2962080_The_Case_for_Energy- Proportional_Computing • Wikipedia https://en.wikipedia.org/wiki/Energy_proportional_computing https://en.wikipedia.org/wiki/Algorithmic_efficiency -
References and further reading Patterns • Green Software Patterns https://patterns.greensoftware.foundation/
• SoftAWERE Principles for Sustainable Software Design https://knowledge.sdialliance.org/softawere/principles-for-sustainable- software-design • Sustainability Pillar – AWS Well-Architected Framework https://docs.aws.amazon.com/wellarchitected/latest/sustainability- pillar/sustainability-pillar.html • Reduce your Google Cloud carbon footprint https://cloud.google.com/architecture/reduce-carbon-footprint -
References and further reading Patterns • Pereira et al., “Energy
Efficiency across Programming Languages: how do energy, time, and memory relate?” https://www.researchgate.net/publication/320436353_Energy_efficiency_ across_programming_languages_how_do_energy_time_and_memory_relat e • Demaine et al. “Energy-Efficient Algorithms” https://arxiv.org/abs/1605.08448 -
References and further reading Guides • Measuring Your Application Power
and Carbon Impact https://devblogs.microsoft.com/sustainable-software/measuring-your- application-power-and-carbon-impact-part-1/ • How to measure the power consumption of your backend service https://devblogs.microsoft.com/sustainable-software/how-to-measure- the-power-consumption-of-your-backend-service/ • How To Measure The Power Consumption of Your Frontend Application https://devblogs.microsoft.com/sustainable-software/how-to-measure- the-power-consumption-of-your-frontend-application/ -
References and further reading Tools • Cloud Carbon Footprint https://www.cloudcarbonfootprint.org/
• Electricity Maps https://www.electricitymaps.com/ • PowerAPI https://github.com/powerapi-ng/powerapi • CodeCarbon https://mlco2.github.io/codecarbon/ -
References and further reading Tools • AWS Customer Carbon Footprint
Tool https://aws.amazon.com/de/aws-cost-management/aws-customer- carbon-footprint-tool/ • Google Cloud Carbon Footprint https://cloud.google.com/carbon-footprint • Emissions Impact Dashboard for Azure https://appsource.microsoft.com/en-us/product/power-bi/coi- sustainability.emissions_impact_dashboard • KEDA – Kubernetes Event-driven Autoscaling https://keda.sh/ -
References and further reading Miscellaneous • Center for Sustainable Systems,
“Green IT fact sheet” https://css.umich.edu/publications/factsheets/built-environment/green-it- factsheet • The Carbon Benefits of Cloud Computing: a Study of the Microsoft Cloud https://www.microsoft.com/en-us/download/details.aspx?id=56950 • 451 Research, “The Carbon Reduction Opportunity of Moving to Amazon Web” https://d39w7f4ix9f5s9.cloudfront.net/e3/79/42bf75c94c279c67d777f00 2051f/carbon-reduction-opportunity-of-moving-to-aws.pdf -
References and further reading Miscellaneous • Center for Sustainable Systems,
“Green IT fact sheet” https://css.umich.edu/publications/factsheets/built-environment/green-it- factsheet • BCS, “Green IT and net zero” https://www.bcs.org/articles-opinion-and-research/green-it/ • Minifesto http://minifesto.org/ • Thompson et al., “Deep Learning’s diminishing Returns” https://spectrum.ieee.org/deep-learning-computational-cost