Building a green supercomputer in the cloud

A case study implementing a carbon aware high-performance compute (HPC) cluster running wind turbine simulations for Vestas. Using the principals of green software and cloud scalability to design a solution that can to scale to 400k cores and multiple petabytes of data. Using open-source tools, UX, and data to lower carbon emissions.

Objectives:

Explain the principals of green software
Apply best practices and tools to reduce carbon reductions
Illustrate the financial impact of carbon emission reductions to the business
Select areas with large carbon impact

Transcript

1. Building a Green Supercomputer in the Cloud Anders Lybecker | @Lybecker
2. Agenda • Introduction • Carbon Efficiency • Energy Efficiency • Hardware Efficiency • Carbon Awareness • Case: Vestas supercomputer • Tools
3. © 2021 Green Software Foundation Why?
4. The IT industry emits more carbon than the aviation industry
5. What can you as a software engineer do?
6. What do we need to do?
8. Power usage effectiveness
9. © 2021 Green Software Foundation Carbon Awareness
10. CO2
11. CO2 eq
12. Carbon intensity
13. CO2 eq/KWH
14. Electricity mix
15. © 2021 Green Software Foundation Electricity mix
19. Vestas
20. Challenges • Wind Turbine Simulation • Optimal placement of turbines in windfarm • Structural optimizations • Existing on-premise HPC cluster • Expensive to maintain • Industry growth at 300% • Complexity of turbine
21. Requirements Destilled Easy of use / same tooling Scale to 400K cores Data segmentation per project Cost management per project Inexpensive As green as possible
22. Design Decisions Saturate CPU Rightsizing VM Auto scale up and down Using unused VMs Time shifting Location shifting
23. DC Client calls API with job spec + files in artifact storage
24. API sends a job create message to the dispatcher queue
25. Dispatcher handles the message
26. Dispatcher selects the pool + copies input data to satellite storage
27. ALT: Dispatcher needs to scale pool and resends the message with a 1 min delay
28. Files are copied back on task complete using event grid (BlobCreated)
29. 30-90% reduction with location shifting with EU 12-16% reduction with time shifting within a day
30. © 2021 Green Software Foundation Tools
31. Carbon Aware SDK
32. Pay 5% extra for 50% carbon reduction
33. Pay 50% extra for 5% carbon reduction
34. Carbon Economy SDK
35. Cost evaluation formular
36. Example Location Emission rating Unit price Norm. rating Norm. price Cost Sweden Central 29 $1.932 0 1 0.750 Norway East 42 $1.824 0.107 0.870 0.653 West Europe 100 $1.103 0.587 0 0.147 UK South 150 $1.713 1 0.736 0.606 Emission weight: 0.25 Location Emission rating Unit price Norm. rating Norm. price Cost Sweden Central 29 $1.932 0 1 0.250 Norway East 42 $1.824 0.107 0.870 0.232 West Europe 100 $1.103 0.587 0 0.440 UK South 150 $1.713 1 0.736 0.772 Emission weight: 0.75
38. Carbon aware KEDA Operator Electricity Carbon Intensity for k8s Region (ConfigMap) Admin 1 Get current Carbon Intensity 2 Get maxReplicasByCarbonIntensity CarbonAwareKedaScaler CRD CarbonAwareKeda Operator scaledObject 3 Take Action: Update maxReplicaCount based on the current Carbon Intensity • Throttle/limit scaling when Carbon Intensity is high • Allow maximum scaling when it is low Every 1 hour: CarbonAwareKeda Operator: 1 Gets the current carbon intensity 2 Gets KEDA targetRef scaling inputs 3 Updates targetRef MaxReplicaCount based on current carbon intensity Step 0: Admin creates CarbonAwareKedaScaler https://github.com/Azure/carbon-aware-keda-operator/blob/main/demo/azure.md
39. © 2021 Green Software Foundation Questions?
43. What makes software green?
45. Environmental impact Potential pathways for improvement Targeted investments Business value
46. Our future, our responsibility
48. require planet-scale innovation and business transformation Planet-sized challenges
49. Microsoft is committed to harnessing the power of technology to help everyone, everywhere build a more sustainable future.
51. Our core environmental sustainability commitments Carbon negative by 2030 Water positive by 2030 Zero waste by 2030 Protect more land than we use and build a Planetary Computer
52. 3%
53. © 2021 Green Software Foundation Measuring: What you can't measure, you can't improve