Green IO Conference at apidays Paris 2024 - The Internet is Getting Emptier; That's a Sustainability Problem, Romain Jacob , ETH Zürich

The Internet is Getting Emptier; That's a Sustainability Problem
Romain Jacob, Senior Researcher at ETH Zürich

Green IO Conference
apidays Paris 2024 - The Future API Stack for Mass Innovation
December 3 - 5, 2024

------

Check out our conferences at https://www.apidays.global/

Do you want to sponsor or talk at one of our conferences?
https://apidays.typeform.com/to/ILJeAaV8

Learn more on APIscene, the global media made by the community for the community:
https://www.apiscene.io

Explore the API ecosystem with the API Landscape:
https://apilandscape.apiscene.io/

Transcript

1. The Internet is getting emptier; that’s a sustainability problem GreenIO - Paris Romain Jacob Dec. 4, 2024 ETH Zürich
2. What do you think consumes more energy? Data Centers Telco Networks
3. What do you think consumes more energy? Data Centers Telco Networks or In 2022 240-340 260-360 TWh TWh https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
4. What do you think consumes more energy? Data Centers Telco Networks or In 2022 240-340 260-360 In 2015 200 220 Change of +20-70% +18-64% TWh TWh in energy TWh TWh in energy https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
5. What do you think consumes more energy? Data Centers Telco Networks or In 2022 240-340 260-360 In 2015 200 220 Change of +20-70% +18-64% +340% +600% TWh TWh in energy in workload TWh TWh in energy in traffic https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks
6. Energy efficiency improved a lot Data Centers Telco Networks Change in energy than in work done. +20-70% +18-64% +340% +600% in energy in workload in energy in traffic ≪
7. Energy efficiency improved a lot but not enough! Data Centers Telco Networks Change in energy +20-70% +18-64% in energy in energy > 0 !
8. With great power comes great responsibility” “  It is easy to keep increasing network capacity  It is much harder to keep increasing energy efficiency
9. With great power comes great responsibility” and carbon footprint. “  It is easy to keep increasing network capacity  It is much harder to keep increasing energy efficiency Producing electricity emits carbon. https://ourworldindata.org/grapher/electricity-prod-source-stacked 60% of the world’s electricity comes from carbon-intensive sources
10. With great power comes great responsibility” and carbon footprint. “  It is easy to keep increasing network capacity  It is much harder to keep increasing energy efficiency Producing electricity emits carbon. Total electricity usage is likely to keep increasing. It doubled in my lifetime.
11. Internet access is still far from universal. https://ourworldindata.org/grapher/share-of-individuals-using-the-internet
12. Let’s zoom in. What’s going on in the Internet?
14. Here is the topology of a middle size Swiss network: Switch. the “map”
15. Average link utilization in the Switch network [%]
16. Average link utilization in the Switch network [%] What do you think is the average link load in Switch?
17. What do you think is the average link load in Switch? 2.1% 2.5 months of data internal links only Average link utilization in the Switch network [%]
18. While we cannot generalize, single digit numbers are not rare. Average link utilization in the Switch network [%] 2.1% 2.5 months of data internal links only
19. The total volume of Internet traffic increases over time. https://www.sciencedirect.com/science/article/pii/S2542435121002117 Traffic Telefonica Energy Cogent [Normalized]
20. But that increase happens only at some times of the day. Telefonica https://eng.ox.ac.uk/media/qhwpxjua/david_sustainability_on_carrier_transport_networks.pdf [Normalized]
21. Today Empty Full
22. As the Internet gets bigger, the difference between “empty” and “full” increases. Today Tomorrow Empty Full
23. Networks are getting emptier, on average. Why does it matter?
24. SIGCOMM 2003 The Internet core consumes more Joules per Bytes than wireless LANs.
25. 2x and 24x more... depending on your hypotheses SIGCOMM 2003 The Internet core consumes more Joules per Bytes than wireless LANs.
26. Network devices are always “on.” 1
27. Network devices are always “on.” Network devices’ energy consumption is mainly independent of traffic load. 1 2
28. The Internet is not really a network of roads Roads do not consume energy when they are unused.
29. The Internet is not really a network of roads; more like Hyperloop tubes. Hyperloop tubes require loads of power to create vacuum, whether used or not. Roads do not consume energy when they are unused.
30. Network devices are always “on.” Network devices’ energy consumption is mainly independent of traffic load. 1 2
31. Network devices are always “on.”  Peak traffic  Fault tolerance ISP overprovision networks to support Network devices’ energy consumption is mainly independent of traffic load. Network devices are under-utilized. 1 2 3
32. Network devices are always “on.”  Peak traffic  Fault tolerance ISP overprovision networks to support Network devices’ energy consumption is mainly independent of traffic load. Network devices are under-utilized. 1 2 3
33. 1 2 3  Peak traffic  Fault tolerance ISP overprovision networks to support Network devices are always “on.” Network devices’ energy consumption is mainly independent of traffic load. Network devices are under-utilized. what you get what you want
34. what you get what you want There two ways to improve energy efficiency  Run more often at high utilization Time-shifting Buffer-and-Burst” “
35. what you get what you want There two ways to improve energy efficiency  Run more often at high utilization Time-shifting Buffer-and-Burst” “
36. what you get what you want There two ways to improve energy efficiency  Run more often at high utilization Time-shifting Buffer-and-Burst”  Take low-utilization power down “
37. what you get what you want There two ways to improve energy efficiency  Run more often at high utilization Time-shifting Buffer-and-Burst”  Take low-utilization power down “
38. The basic idea is to turn off “stuff” whenever possible. Router
39. The basic idea is to turn off “stuff” whenever possible. Router Transceiver
40. The basic idea is to turn off “stuff” whenever possible.  Ports  Line cards  Entire device... What can we possibly turn off? Router Transceiver
41. The basic idea is to turn off “stuff” whenever possible.  Ports  Line cards  Entire device...  Memory banks  Power supplies  LEDs ... etc. What can we possibly turn off? Router Transceiver
42. The basic idea is to turn off “stuff” whenever possible.  Ports  Line cards  Entire device...  Memory banks  Power supplies  LEDs ... etc. It can be more subtle than on/off.  Change a port rate from 100G to 10G  Down-clock the ASIC  Cache frequently used FIB entries What can we possibly turn off?
43. The basic idea is to turn off “stuff” whenever possible. That’s nothing new. RIPE Academia NSDI 2008 86
44. The theory says we can save tens of energy % in ISP networks. Energy Savings (%) Academia NSDI 2008
45. The theory says we can save tens of energy % in ISP networks. Academia Buffer-and-Burst How? Assumes  Wake-up delay 1𝑚𝑠  Buffering time 10𝑚𝑠 NSDI 2008
46. Buffer-and-Burst How? Assumes  Wake-up delay 1𝑚𝑠  Buffering time 10𝑚𝑠 Theory
47. Buffer-and-Burst How? Assumes  Wake-up delay 1𝑚𝑠  Buffering time 10𝑚𝑠 Wake-up delay Measured on Cisco Nexus 9300 (s) Theory Practice VS
48. Buffer-and-Burst How? Wake-up delay Assumes  Wake-up delay 1𝑚𝑠  Buffering time 10𝑚𝑠 Measured on Cisco Nexus 9300 In practice, transcievers are 1000x slower to start than required for savings via buffering (today). (s)
49. We can still “sleep” at longer timescales. That is not very different from other routine network management tasks. Average link utilization in the Switch network [%]
50. On Switch, Hypnos can turn off ~1/3 of the links. Hypnos – Greek god of sleep  Presented at HotCarbon’24  Paper and presentation  Evaluation code on GitHub  Hardware implementation for an earlier prototype
51. How much energy can we really save? With Hypnos?
52. How much energy can we really save? Very little. (<1%) With Hypnos?
53. How much energy can we really save? Very little. (<1%) Because “turned off” does not always mean “powered off.” Hyperloop tubes require loads of power to create vacuum, whether used or not.
54. Router Transceiver QSFP28 LR EdgeCore Wedge 100BF-32X
55. Router Transceiver 3-4 W 110 W “empty” per piece QSFP28 LR EdgeCore Wedge 100BF-32X
56. We are doing a terrible job at power proportionality today. Router Transceiver 3-4 W 110 W “empty” per piece QSFP28 LR EdgeCore Wedge 100BF-32X Power increases at plug-in even if the port is “off”!
57. We are doing a terrible job at power proportionality today. https://community.juniper.net/blogs/nicolas-fevrier/2023/09/11/saving-power-on-acx7000-series Similar problems with internal router components that remain powered on while being configured “off.”  Thankfully, hardware vendors are starting to notice...  ... and it is actually easy to fix!
58. We are doing a terrible job at power proportionality today. It’s a bug. https://community.juniper.net/blogs/nicolas-fevrier/2023/09/11/saving-power-on-acx7000-series Similar problems with internal router components that remain powered on while being configured “off.”  Thankfully, hardware vendors are starting to notice...  ... and it is actually easy to fix! Until this gap is addressed, a simple workaround exists: configuring the empty ports in “unused” mode. In case you wonder The usual config is “shutdown.”
59. Don’t fill the pipes! Why we must keep the Internet empty.
60. Once the network is up, traffic costs almost nothing. Average link utilization in the Switch network [%] 1 2 [Tbps]
61. Once the network is up, traffic costs almost nothing. Average link utilization in the Switch network [%] The power to send that traffic is 1 2 [Tbps] That’s 0.02% of the total network power. ~6W
62. what you get what you want Increasing utilization improves efficiency  Run more often at high utilization  Take low-utilization power down
63. what you get what you want Increasing utilization improves efficiency  Run more often at high utilization  Take low-utilization power down in the short term.
64. what you get what you want Increasing utilization improves efficiency  Run more often at high utilization  Take low-utilization power down in the short term.  Peak traffic  Fault tolerance ISP overprovision networks to support Do not forget
65. https://doi.org/10.1145/2858036.2858378
66. https://doi.org/10.1145/2858036.2858378
67. https://doi.org/10.1145/2858036.2858378
68. There is a feedback loop that stimulates network capacity increase https://doi.org/10.1145/2858036.2858378
69. There is a feedback loop that stimulates network capacity increase and energy usage. https://research-information.bris.ac.uk/en/publications/rethinking- allocation-in-high-baseload-systems-a-demand-proportio https://doi.org/10.1145/2858036.2858378
70. As “we” keep asking for more, the energy use will keep rising.
71. We must embrace some digital sobriety. Everything has a cost.  Every picture we upload  Every conversation we archive  Every app we download  Every movie we stream
72. We must embrace some digital sobriety. Everything has a cost.  Every picture we upload  Every conversation we archive  Every app we download  Every movie we stream It is not to say we must not do it we must be mindful when doing it, and do it only when actually useful. but
73. We must embrace some digital sobriety. Everything has a cost. It is not to say we must not do it we must be mindful when doing it, and do it only when actually useful. but That does not have to imply worse UX!  Every picture we upload  Every conversation we archive  Every app we download  Every movie we stream  Software bloat can be huge.  Applications can be redesigned without assuming permanent high-speed connectivity. That’s where GreenIO folks come in!
74. ... and what’s with the Zoo? What is the most surprising word that you will use in your speech? Zoo
75. We have a data transparency problem for networking hardware.  Vendors hardly provide any useful power info.  Embedded power sensors are not trustworthy.  There is no standard for power data reporting. TL;DR We need more power measurements and a systematic data collection.
76. We created a public database for power data: Network Power Zoo  Datasheet information  PSU readings  External measurements  Power models Would you share your network’s data? We work on tools to make it easy The database contains networkpowerzoo.ethz.ch Welcome to the
78. 1 The Internet peak traffic increases. That is driving network growth, which results in emptier networks on average.
79. 1 The Internet peak traffic increases. That is driving network growth, which results in emptier networks on average. Networking hardware is far from power proportionality; It is very inefficient at low utilization. 2 It can and must be improved.
80. 1 The Internet peak traffic increases. That is driving network growth, which results in emptier networks on average. 3 Even if direct cost of traffic is minuscule, sending more traffic has an important systemic cost. Cf. Point 1. Networking hardware is far from power proportionality; It is very inefficient at low utilization. 2 Don’t send if you don’t need to. Stay away from the peak! It can and must be improved.
81. jacobr@ethz.ch Romain Jacob An Hyperloop tube segment. The Internet is getting emptier; that’s a sustainability problem that we can fix.