The Efficiency Paradox and How to Save Yourself and the World

Inefficiency is ruining our planet and our lives. Efficiency is ruining our happiness, and weirdly, it’s also ruining our efficiency. Heeeeelppp!? What’s a techie to do? In this talk, Holly walks through a range of techniques that can be used to find and eliminate software waste and reduce climate impacts. For example, zombie servers and slow code are a big climate problem, but the vrrrooom model gives us double-win hope.

But machine efficiency isn’t much use without human efficiency. The good news is that there isn’t as much of a tradeoff as you might think between these two types of efficiency. Sorting out machine efficiency often helps humans, too. For example, the Quarkus Java framework uses many interesting waste-reduction techniques. These optimisations have the dual benefit of speeding up computers, and also speeding up people.

So far, so good, but we need to be careful we don’t end up accidentally optimising the wrong things, and making stuff worse. 100% utilisation is not sustainable for either humans, or people. It’s not even very efficient (remember the paradox part?). It turns out, thinking about efficiency needs systems thinking. Holly will draw lessons from performance engineering, mathematics,economics, and psychology to give guidance on how to achieve more by doing less.

Transcript

1. @hollycummins.com QCon London Holly Cummins Red Hat The Efficiency Paradox How to Save Yourself … and the World
2. @holly_cummins #RedHat now senior principal software engineer helping to build Quarkus
3. @holly_cummins #RedHat now senior principal software engineer helping to build Quarkus 2007 let’s make garbage collection more efficient!
4. @holly_cummins #RedHat now senior principal software engineer helping to build Quarkus 2007 let’s make garbage collection more efficient! 2015 lean and xp makes your team more efficient!
5. @holly_cummins #RedHat now senior principal software engineer helping to build Quarkus 2007 let’s make garbage collection more efficient! 2015 lean and xp makes your team more efficient! 2022 quarkus is wonderfully efficient!
6. @holly_cummins #RedHat “this provisioning software is broken”
7. @holly_cummins #RedHat what we sold “this provisioning software is broken” 10 minute provision-time
8. @holly_cummins #RedHat what we sold “this provisioning software is broken” 10 minute provision-time 3 month provision- time what the client thought they’d got
9. @holly_cummins #RedHat what we sold “this provisioning software is broken” 10 minute provision-time 3 month provision- time what the client thought they’d got the reason 84-step pre-approval process
10. @holly_cummins #RedHat waste
11. @hollycummins.com 1700s machine
12. @hollycummins.com 1700s machine energy
13. @hollycummins.com 1700s machine energy energy
14. @hollycummins.com 1700s machine energy energy useful
15. @hollycummins.com 1700s waste machine energy energy useful
16. @hollycummins.com 1890s processes
17. @hollycummins.com time money 1890s processes
18. @hollycummins.com time money 1890s value processes
19. @hollycummins.com time money 1890s value processes
20. @hollycummins.com time money 1890s value processes waste
21. @hollycummins.com software 1960s
22. @hollycummins.com software time electricity hardware 1960s
23. @hollycummins.com software time electricity hardware answers 1960s
24. @hollycummins.com software time electricity hardware answers 1960s waste
25. @hollycummins.com waste is killing the planet
26. @hollycummins.com e-waste is killing the planet
27. @hollycummins.com energy waste is killing the planet
28. @hollycummins.com slow code is killing the planet
29. @hollycummins.com https://blog.linkedin.com/2017/august/3/making-linkedin-more-accessible-via-linkedin-lite Nashik
30. @hollycummins.com https://blog.linkedin.com/2017/august/3/making-linkedin-more-accessible-via-linkedin-lite “my heart sank … our new feature failed to load because of poor internet connectivity” Nashik
31. @hollycummins.com https://blog.linkedin.com/2017/august/3/making-linkedin-more-accessible-via-linkedin-lite modern web is so inefficient it is useless for part of its audience “my heart sank … our new feature failed to load because of poor internet connectivity” Nashik
32. @hollycummins.com forgotten code is killing the planet
33. #RedHat @holly_cummins@hachyderm.io 25% of 16,000 servers doing no useful work https://www.anthesisgroup.com/wp-content/uploads/2019/11/Comatose-Servers-Redux-2017.pdf zombie servers
34. #RedHat @holly_cummins@hachyderm.io the average server: 12 - 18% of capacity https://www.nrdc.org/sites/default/files/data-center-efficiency-assessment-IB.pdf
35. @hollycummins.com unused data is killing the planet
36. #RedHat @holly_cummins@hachyderm.io 68% of data is stored and never used https://www.nrdc.org/sites/default/files/data-center-efficiency-assessment-IB.pdf https://www.seagate.com/gb/en/news/news-archive/seagates-rethink-data-report-reveals-that-68-percent-of-data-available-to-businesses-goes-unleveraged-pr-master/
37. @hollycummins.com it’s not just money
38. @hollycummins.com it’s not just electricity it’s not just money
39. @hollycummins.com it’s not just electricity it’s not just money it’s embodied carbon
40. @hollycummins.com it’s not just electricity it’s water it’s not just money it’s embodied carbon
41. @hollycummins.com it’s not just electricity it’s water it’s e-waste it’s not just money it’s embodied carbon
42. @hollycummins.com we have solutions good news, everyone
43. @holly_cummins #RedHat making turning servers off as safe and easy as turning lights off. and then automate it.
44. @holly_cummins #RedHat solution: LightSwitchOps making turning servers off as safe and easy as turning lights off. and then automate it.
45. @holly_cummins #RedHat absurdly simple scripts my shell script to power down machines overnight saved my school €12,000
46. @holly_cummins #RedHat simple automation we used to leave our applications running all the time @darkandnerdy, Chicago DevOpsDays
47. @holly_cummins #RedHat simple automation we used to leave our applications running all the time when we scripted turning them off at night, we reduced our cloud bill by 30% @darkandnerdy, Chicago DevOpsDays
48. @holly_cummins #RedHat fancier scripts … with a frontend and a backend
49. @holly_cummins #RedHat
50. @holly_cummins #RedHat saves energy
51. @holly_cummins #RedHat saves energy saves money
52. @holly_cummins #RedHat saves energy saves money bonus: tests disaster recovery
53. @holly_cummins #RedHat saves energy saves money bonus: tests disaster recovery
54. @holly_cummins #RedHat
55. @holly_cummins #RedHat solution: faster code
56. #Quarkus @holly_cummins@hachyderm.io focus on making the bottleneck efficient
57. @holly_cummins the vrroooom model
58. @holly_cummins
59. @holly_cummins invented by Dr. Vroom (really!)
60. @holly_cummins naming is the hardest problem in computer science invented by Dr. Vroom (really!)
61. @holly_cummins naming is the hardest problem in computer science
62. @holly_cummins my vrroooom model
64. @holly_cummins #RedHat
65. @holly_cummins #RedHat these two columns are almost the same
66. @holly_cummins #RedHat Energy 1 10 100 Time 1 10 100 energy efficiency across programming languages Python Rust Java Go
67. @holly_cummins #RedHat Energy 1 10 100 Time 1 10 100 the trend line is more or less straight energy efficiency across programming languages Python Rust Java Go
69. case study Quarkus
70. @holly_cummins #RedHat Java application frameworks needed dynamic dependencies the distant past
71. @holly_cummins #RedHat cloud apps are immutable now
72. @holly_cummins #RedHat cloud apps are immutable now
73. @holly_cummins #RedHat a highly dynamic runtime in a container is pointless
74. @holly_cummins Java dynamism
75. @holly_cummins Java dynamism build time
76. @holly_cummins Java dynamism build time runtime
77. @holly_cummins Java dynamism build time runtime
78. @holly_cummins Java dynamism packaging (maven, gradle…) build time runtime
79. @holly_cummins Java dynamism build time runtime
80. @holly_cummins Java dynamism build time runtime
81. @holly_cummins Java dynamism </> build time runtime load and parse • config files • properties • yaml • xml • etc.
82. @holly_cummins Java dynamism </> build time runtime
83. @holly_cummins Java dynamism @ @ </> build time runtime • classpath scanning and annotation discovery • attempt to load class to enable/disable features
84. @holly_cummins Java dynamism @ @ </> build time runtime
85. @holly_cummins Java dynamism @ @ </> build time runtime build a metamodel of the world
86. @holly_cummins Java dynamism @ @ </> build time runtime
87. @holly_cummins Java dynamism @ @ </> build time runtime start • thread pools • I/O • etc.
88. @holly_cummins Java dynamism @ @ </> build time runtime ready to do work!
89. @holly_cummins what if we start the application more than once? @ @ </>
90. @holly_cummins what if we start the application more than once? @ @ </> @ @ </>
91. @holly_cummins what if we start the application more than once? @ @ </> @ @ </> @ @ </>
92. @holly_cummins what if we start the application more than once? @ @ </> @ @ </> @ @ </> @ @ </>
93. @holly_cummins what if we start the application more than once? @ @ </> @ @ </> @ @ </> @ @ </>
94. @holly_cummins what if we start the application more than once? @ @ </> @ @ </> @ @ </> @ @ </> so much work gets redone every time
95. @holly_cummins Hibernate speed example: JTA auto-wiring
96. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”);
97. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”);
98. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”);
99. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”);
100. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); …
101. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); Class.forName(“NicheJTAImplementation”); …
102. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); Class.forName(“NicheJTAImplementation”); Class.forName(“VeryNicheJTAImplementation”); …
103. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); Class.forName(“NicheJTAImplementation”); Class.forName(“VeryNicheJTAImplementation”); …
104. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); Class.forName(“NicheJTAImplementation”); Class.forName(“VeryNicheJTAImplementation”); … ~129 auto-wiring attempts
105. @holly_cummins Hibernate speed example: JTA auto-wiring Class.forName(“LikelyJTAImplementation”); Class.forName(“APossibleJTAImplementation”); Class.forName(“AnotherJTAImplementation”); Class.forName(“NicheJTAImplementation”); Class.forName(“VeryNicheJTAImplementation”); … ~129 auto-wiring attempts every single start.
106. @holly_cummins the true cost of loaded classes isn’t just memory + start time
107. @holly_cummins the true cost of loaded classes isn’t just memory + start time method dispatching:
108. @holly_cummins interface the true cost of loaded classes isn’t just memory + start time method dispatching:
109. @holly_cummins unused implementation the one we want interface unused implementation unused implementation the true cost of loaded classes isn’t just memory + start time method dispatching:
110. @holly_cummins unused implementation the one we want interface unused implementation unused implementation the true cost of loaded classes isn’t just memory + start time method dispatching:
111. @holly_cummins unused implementation the one we want interface megamorphic call slow dispatching unused implementation unused implementation the true cost of loaded classes isn’t just memory + start time method dispatching:
112. @holly_cummins the true cost of loaded classes isn’t just memory + start time the one we want interface
113. @holly_cummins the true cost of loaded classes isn’t just memory + start time the one we want monomorphic call fast dispatching interface
114. @holly_cummins how do we fix all this?
115. @holly_cummins @ @ </> build time runtime what if we initialize at build time?
116. @holly_cummins @ @ </> build time runtime what if we initialize at build time?
117. @holly_cummins @ @ </> build time runtime ready to do work! what if we initialize at build time?
118. @holly_cummins @ @ </> build time runtime ready to do work! start • thread pools • I/O • etc. what if we initialize at build time?
119. @holly_cummins @ @ </> repeated starts are now efficient
120. @holly_cummins @ @ </> repeated starts are now efficient
121. @holly_cummins @ @ </> repeated starts are now efficient
122. @holly_cummins @ @ </> repeated starts are now efficient
123. @holly_cummins @ @ </> repeated starts are now efficient
124. @holly_cummins @ @ </> repeated starts are now efficient
125. @holly_cummins @ @ </> repeated starts are now efficient less wasted work
126. @holly_cummins #RedHat capacity Source: John O’Hara Setup: • REST + CRUD • large heap • RAPL energy measurement Assumptions: • US energy mix climate impact of framework choice
127. @holly_cummins #RedHat capacity Source: John O’Hara Setup: • REST + CRUD • large heap • RAPL energy measurement Assumptions: • US energy mix climate impact of framework choice shorter line means lower max throughput
128. @holly_cummins #RedHat capacity Source: John O’Hara Setup: • REST + CRUD • large heap • RAPL energy measurement Assumptions: • US energy mix climate impact of framework choice shorter line means lower max throughput higher line means worse carbon footprint
129. @holly_cummins #RedHat capacity Source: John O’Hara Setup: • REST + CRUD • large heap • RAPL energy measurement Assumptions: • US energy mix climate impact of framework choice vrrrooom model in action: quarkus on JVM has the smallest footprint … because it has the highest throughput shorter line means lower max throughput higher line means worse carbon footprint
130. @holly_cummins throughput startup time + footprint
131. @holly_cummins we beat the trade-off. throughput startup time + footprint
132. @holly_cummins we beat the trade-off. throughput startup time + footprint
133. @holly_cummins we beat the trade-off. throughput startup time + footprint it’s a double-win.
135. @holly_cummins we beat another trade-off. human efficiency machine efficiency
136. @holly_cummins we beat another trade-off. human efficiency machine efficiency
137. @holly_cummins we beat another trade-off. human efficiency machine efficiency another double-win.
138. @hollycummins.com boilerplate code
139. @hollycummins.com boilerplate code
140. @hollycummins.com live coding
141. @hollycummins.com live coding -less energy wasted running full build + launch cycles
142. @hollycummins.com live coding -less energy wasted running full build + launch cycles -less dev time wasted waiting for full build + launch cycles
143. @hollycummins.com targeted continuous testing
144. @hollycummins.com targeted continuous testing -less energy wasted running full test suites
145. @hollycummins.com targeted continuous testing -less energy wasted running full test suites -less dev time wasted waiting for full test runs
146. @holly_cummins developer joy
147. @holly_cummins developer joy
148. @holly_cummins developer joy is a double win
149. @holly_cummins the double win isn’t just a java thing.
150. @holly_cummins the ai-lephant in the room
151. @holly_cummins we don’t need to have a nice programming model; people can just use AI to write the bad code!
152. @holly_cummins we don’t need to have a nice programming model; people can just use AI to write the bad code! 3 problems.
153. @holly_cummins problem 1: cost
154. @holly_cummins problem 1: cost
155. @holly_cummins #RedHat problem 2: correctness
156. @holly_cummins #RedHat efficiency of kangaroo- powered renewable energy problem 2: correctness
157. @holly_cummins #RedHat efficiency of kangaroo- powered renewable energy 15% problem 2: correctness
158. @holly_cummins #RedHat
159. @holly_cummins #RedHat
160. @holly_cummins #RedHat
161. @holly_cummins #RedHat disclaimer:
162. @holly_cummins #RedHat disclaimer: this isn’t a thing.
163. @holly_cummins #RedHat disclaimer: this isn’t a thing. researchers did not use kangaroos on trampolines for electricity generation.
164. @hollycummins.com problem 3: verbosity
165. @hollycummins.com problem 3: verbosity
166. @hollycummins.com problem 3: verbosity
167. @hollycummins.com problem 3: verbosity
168. @hollycummins.com problem 3: verbosity
169. @hollycummins.com problem 3: verbosity
170. @hollycummins.com 70% unnecessary code problem 3: verbosity
171. @hollycummins.com 70% more code than we need hurts us
172. @hollycummins.com 70% more code than we need hurts us we spend far more time reading code than writing it
173. @hollycummins.com “illusion of efficiency”
174. @hollycummins.com we have a solution
175. @holly_cummins -small fine-tuned models -small model + RAG -hybrid: symbolic reasoning + model solution:
176. @holly_cummins illusions of efficiency are everywhere
177. @hollycummins.com limits to efficiency
178. @holly_cummins #RedHat Jevon’s paradox limit 1: economic theory “efficiency improvements can sometimes lead to increased consumption”
179. @holly_cummins #RedHat what we imagine when we widen roads
180. @holly_cummins #RedHat what we get
181. @holly_cummins #RedHat limit 2: physics
182. @holly_cummins #RedHat even for machines, there is a limit to efficiency limit 2: physics
183. @holly_cummins #RedHat even for machines, there is a limit to efficiency …and it’s lower than you think limit 2: physics
184. @hollycummins.com source: http://geosci.uchicago.edu/~moyer/GEOS24705/2016/Assignments/PS7.pdf physics limit 2
185. @holly_cummins #RedHat limit 2
186. @holly_cummins #RedHat limit 2
187. @holly_cummins #RedHat theoretical max efficiency of a combustion engine limit 2
188. @holly_cummins #RedHat theoretical max efficiency of a combustion engine limit 2 37%
189. @hollycummins.com efficiency can give worse outcomes
190. @holly_cummins #RedHat typical efficiency of a combustion engine
191. @holly_cummins #RedHat ~20% typical efficiency of a combustion engine
192. @holly_cummins #RedHat if we want the engine to not wear out ~20% typical efficiency of a combustion engine
193. @holly_cummins resiliency efficiency
194. @hollycummins.com
195. @hollycummins.com speed
196. @hollycummins.com speed vs resiliency
197. @hollycummins.com let’s talk about your multi-region failover.
198. @hollycummins.com #RedHat
199. @hollycummins.com highly efficient (optimum number of legs) #RedHat
200. @hollycummins.com highly efficient (optimum number of legs) less efficient (more legs than needed) #RedHat
201. @hollycummins.com highly efficient (optimum number of legs) less efficient (more legs than needed) resilient #RedHat
202. @hollycummins.com highly efficient (optimum number of legs) less efficient (more legs than needed) resilient #RedHat
203. @hollycummins.com highly efficient (optimum number of legs) no resiliency less efficient (more legs than needed) resilient #RedHat
204. @hollycummins.com highly efficient (optimum number of legs) no resiliency less efficient (more legs than needed) resilient #RedHat resiliency lowers efficiency
205. @hollycummins.com it’s the same for people
206. @hollycummins.com it’s the same for people all work and no play is … not efficient
207. @hollycummins.com
208. @hollycummins.com
209. @hollycummins.com
210. @hollycummins.com hi colleague, could you ple-
211. @hollycummins.com hi colleague, could you ple- argh! &!*$*%@!{*%^!^! busy! NO! I CANNOT!
212. @hollycummins.com
213. @holly_cummins the value of slack
214. @holly_cummins the value of slack
215. @holly_cummins the value of slack
216. @holly_cummins the value of slack naming is still the hardest problem in computer science
217. @holly_cummins the value of slack
218. @holly_cummins #RedHat doing nothing
219. @holly_cummins #RedHat doing nothing efficiency
220. @holly_cummins #RedHat the default mode network idle minds can solve hard problems
221. #RedHat @holly_cummins
222. #RedHat @holly_cummins
223. #RedHat @holly_cummins
224. #RedHat @holly_cummins
225. #RedHat @holly_cummins 14% took showers specifically for the purpose of coming up with ideas
226. @hollycummins.com psychology says we need idle time; maths agrees
227. @hollycummins.com queueing theory says systems need to run under-capacity to function psychology says we need idle time; maths agrees
228. @hollycummins.com queuing theory basics
229. @hollycummins.com queuing theory basics arrival process
230. @hollycummins.com queuing theory basics arrival process Poisson distribution
231. @hollycummins.com queuing theory basics queue arrival process Poisson distribution
232. @hollycummins.com queuing theory basics queue arrival process Poisson distribution
233. @hollycummins.com queuing theory basics queue arrival process servers Poisson distribution
234. @hollycummins.com queuing theory basics queue arrival process servers Poisson distribution
235. @hollycummins.com queuing theory basics queue arrival process servers completed work Poisson distribution
236. @hollycummins.com if arrival rates are low, servers will be idle queue arrival process servers completed work
237. @hollycummins.com queue arrival process servers completed work if server capacity is too low, wait times are high
238. @hollycummins.com utilisation lead time http://brodzinski.com/2015/01/slack-time-value.html assuming Poisson distribution of arrivals
239. @hollycummins.com utilisation lead time http://brodzinski.com/2015/01/slack-time-value.html assuming Poisson distribution of arrivals 80% utilisation → 90% utilisation: wait times double
240. @hollycummins.com utilisation http://brodzinski.com/2015/01/slack-time-value.html cost delay cost assuming Poisson distribution of arrivals 80% utilisation → 90% utilisation: wait times double
241. @hollycummins.com utilisation http://brodzinski.com/2015/01/slack-time-value.html cost delay cost cost of idle capacity assuming Poisson distribution of arrivals 80% utilisation → 90% utilisation: wait times double
242. @holly_cummins #RedHat the 6-day working week was the standard
243. @holly_cummins #RedHat the 6-day working week
244. @holly_cummins #RedHat
245. @holly_cummins #RedHat productivity stayed the same
246. @holly_cummins #RedHat
247. @holly_cummins #RedHat the 4-day working week?
248. @holly_cummins #RedHat 42% decrease in resignations the 4-day working week?
249. @holly_cummins #RedHat 42% decrease in resignations 36% increase in revenue the 4-day working week?
250. @holly_cummins #RedHat efficiency may not look how we expect
251. @holly_cummins #RedHat most efficient land animal efficiency may not look how we expect
252. @holly_cummins #RedHat most efficient land animal cool and bouncy efficiency may not look how we expect
253. @holly_cummins #RedHat most efficient land animal cool and bouncy uncool and floppy efficiency may not look how we expect
254. @holly_cummins #RedHat most efficient land animal cool and bouncy uncool and floppy whole body is basically slack efficiency may not look how we expect
255. @holly_cummins #RedHat most efficient land animal cool and bouncy uncool and floppy whole body is basically slack most efficient animal efficiency may not look how we expect
256. @holly_cummins #RedHat inverse Jevon’s manoeuvre doing less achieves more
257. @holly_cummins being effective working less
258. @holly_cummins being effective working less
259. @holly_cummins being effective working less another double-win.
260. @hollycummins.com tl;dpa (too long; didn’t pay attention)
261. @hollycummins.com tl;dpa ⁃ don’t accept waste; our world deserves better (too long; didn’t pay attention)
262. @hollycummins.com tl;dpa ⁃ don’t accept waste; our world deserves better ⁃ work less; achieve more (too long; didn’t pay attention)
263. @hollycummins.com tl;dpa ⁃ don’t accept waste; our world deserves better ⁃ work less; achieve more ⁃ happiness is not waste (too long; didn’t pay attention)
264. @hollycummins.com tl;dpa ⁃ don’t accept waste; our world deserves better ⁃ work less; achieve more ⁃ happiness is not waste ⁃ idleness is not waste (too long; didn’t pay attention)
265. @hollycummins.com tl;dpa ⁃ don’t accept waste; our world deserves better ⁃ work less; achieve more ⁃ happiness is not waste ⁃ idleness is not waste ⁃ look for double-wins everywhere (too long; didn’t pay attention)
266. @hollycummins.com slides