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 ...

Transcript

1. Patterns of sustainability Going green in IT Uwe Friedrichsen – codecentric AG – 2020-2023
2. Uwe Friedrichsen Works @ codecentric https://twitter.com/ufried https://www.speakerdeck.com/ufried https://ufried.com/
3. Is green IT a relevant topic?
4. “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)
5. “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)
6. But Green IT is a problem of hardware manufacturers and datacenter designers and operators only, isn't it?
7. 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! 🥳
8. “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)
9. “Software efficiency halves every 18 months, compensating Moore's Law” (May's law) Source: Wikipedia, Algorithmic efficiency (https://en.wikipedia.org/wiki/Algorithmic_efficiency)
10. 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! 🥳
11. What can we as software engineers do to make IT greener?
12. Well, a lot ...
13. Patterns of sustainability (Software engineering perspective) Requirements Hardware manufacturing & data center design
14. 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
15. 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
16. 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
17. 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
18. 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, ...
19. 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
20. 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
21. 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
22. 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
23. 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
24. 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
25. 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
26. 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
27. 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
28. 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
29. 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
30. 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
31. 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
32. 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
33. 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
34. 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
35. 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
36. 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
37. 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
38. 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
39. 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
40. 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
41. 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
42. 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
43. 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
44. 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
45. 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
46. 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
47. 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
48. 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
49. 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
50. 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
51. 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
52. 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
53. 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)
54. 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
55. “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
56. 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/
57. 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
58. Programming languages differ widely with respect to energy, time and memory consumption
59. 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
60. 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
61. 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
62. 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
63. 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
64. 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
65. “Software efficiency halves every 18 months, compensating Moore's Law” (May's law) Source: Wikipedia, Algorithmic efficiency (https://en.wikipedia.org/wiki/Algorithmic_efficiency)
66. “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)
67. 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)
68. 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
69. 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
70. 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
71. 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
72. 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
73. 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/
74. 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
75. 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
76. 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
77. 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
78. 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
79. 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
80. 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
81. 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
82. 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
83. 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
84. 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
85. 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
86. Source: Barroso et al., The Case for Energy-Proportional Computing, IEEE Computer Dec 2007, pp. 33-37, vol. 40
87. 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
88. 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
89. 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
90. 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
91. 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
92. 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
93. 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
94. 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
95. 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
96. 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
97. 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
98. 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
99. 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
100. 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
101. 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
102. 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
103. 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
104. 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
105. 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
106. Source: https://app.electricitymaps.com/map
107. 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
108. 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
109. On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need
110. On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need
111. 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
112. On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need
113. “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
114. 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
115. On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need
116. 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
117. On-premises (Responsible: Operations) Buy energy-efficient hardware Shut down unneeded hardware, infrastructure and applications Buy only the hardware you need
118. 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
119. 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
120. 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
121. 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
122. 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
123. 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
124. 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
125. 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
126. 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
127. 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
128. 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
129. 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
130. 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
131. 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
132. Cross-cutting (Responsible: Everyone) Continuously improve Make energy-consumption visible Do not feed the resource troll
133. Cross-cutting (Responsible: Everyone) Continuously improve Make energy-consumption visible Do not feed the resource troll
134. 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
135. Cross-cutting (Responsible: Everyone) Continuously improve Make energy-consumption visible Do not feed the resource troll
136. 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
137. Cross-cutting (Responsible: Everyone) Continuously improve Make energy-consumption visible Do not feed the resource troll
138. 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
139. Cross-cutting (Responsible: Everyone) Continuously improve Make energy-consumption visible Do not feed the resource troll
140. 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
141. 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
142. The bigger picture
143. Sustainability is more than just ecological sustainability
144. 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
145. Sustainability lives in a bigger context
146. Core IT challenges of the 21st century (Responsible: Everyone) Resilience Sustainability Simplification
147. Summing up
148. 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
149. A journey of a thousand miles begins with a single step -- Chinese proverb
150. 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
151. 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
152. 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
153. 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
154. 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/
155. 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/
156. 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/
157. 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
158. 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
159. Uwe Friedrichsen Works @ codecentric https://twitter.com/ufried https://www.speakerdeck.com/ufried https://ufried.com/