WAD 2022 - 5 ways to optimize cloud costs

In this talk, we presented 5 different ways on how we optimized our cloud costs.

Transcript

1. © 2022 . Brantner Digital Solutions GmbH . Dr.-Franz-Wilhelm-Straße 2a . 3500 Krems . Österreich 1 5 ways to optimize cloud costs V1.1-2022
2. 2 Storage Scalability Containers Queue system Costs as First Class Metric What is this talk about BTW: we like clouds ;)
3. 3 Timetable Interactive Presentation Q&A, Discussions Trying out ♥ We all ♥ to visit many talks. We try to finish early so that you can get the most out of the conference experience.
4. 4 Speaker Dr. Rene Heinzl - Since 2021: Founder, CEO and Managing Partner of BDS - 2015 - 2021: CIO of MTB (Dr. Taus) Industry conglomerate in Austria, 1 billion revenue, 6.000 employees - 2013 - 2015: CEO of amanomedia Part of Leykam, one of the largest printing facilities in Europe - Since 2007: Founder, CEO and Managing Partner of DHC Technology consulting for companies in Europe - 2004-2007: Assistant professor at TU Vienna and visiting researcher at: Intel USA, Portland, Oregon: Applied semiconductor research Sony Japan, Tokyo: Semiconductor research CEO of Brantner Digital Solutions
5. 5 Speaker DI Markus Nissl - Since 2021: - Leading the architecture development of Brantner Digital Solutions - Since 2019: - Project assistant at TU Wien - Researching on Financial, Blockchain and Temporal Knowledge Graphs - International Collaborations - Founded his own company with 18 Architecture Lead
6. 6 Company Information Parent company: Brantner group since 1936 more than 2.700 employees 12 waste sorting facilities Company name: Brantner Digital Solutions GmbH Slogan: Brantner AI - so easy to apply! Specialization: applied artificial intelligence for sustainability Founders: Bernd Brantner, René Heinzl CEO: Christoph Pasching (sales) René Heinzl (tech) Josef Scheidl (finance) Innovation from tradition OFFICES AUSTRIA since 1936 CZECH REPUBLIC | SLOVAKIA since 1992 ROMANIA since 2004 SERBIA since 2007
7. Technology: AI - as - a - service Technology: freely programmable neuronal net (PyTorch and Tensorflow) Training data: 1 20 object categories, approx. 3.9 million images per month Patents: currently 5 applications Our awards (in German): Platform: scalable Kubernetes cluster for on-premise, private and public cloud (we are looking for experienced developers) Neuronal net: panoptic segmentation for object detection, instance segmentation, semantic segmentation 7 Multiple award winning product Presentation of the Brantner AI
8. 8 B-AI: Reduction of plastic waste
9. 9 Our gauge-process
10. 10 Our USP: specialized in fully deformed objects Wir erkennen auch stark deformierte Objekte eindeutig
11. 11 How can we achieve 100 EUR cloud costs per month? Metrics: 200 API calls per second / storage 40 TB / 4 million JSON DB entries per month Storage Scalability Containers Queue system Costs as First Class Metric
12. 12 Storage Scalability Containers Queue system Costs as First Class Metric
13. 13 Make cost a first class metric There are several key metrics that pertain to cloud costs: Measure per Promote the idea of cloud costs Customer Product Feature in the organization understand actual costs Dev Team Environment top-of-mind for engineering decisions Ensure that costs are measureable up-to-date viewed within relevant context
14. 14 Make cost a first class metric There are several key metrics that pertain to cloud costs: Unit cost Idle cost per instance/per call? what factors affect it impact on bottom line Return on unit cost has to be strong baseline cost with 0 load efficiency measurement Helps to determine whether an architecture change is worth the effort in terms of savings you can realize Shared infrastructure save costs or engineering efficiencies chargeback or account for shared costs split costs across teams is often challenging
15. 15 Make cost a first class metric There are several key metrics that pertain to cloud costs: Cost/Load efficiency curve Innovation/cost ratio calculate your costs unit cost vs. customer base avoid exponential growth in your cost curve R&D to production operation costs no expectation of revenue in R&D keep system profitable products will eventually go to market If costs were not considered during R&D, moving to production can be challenging
16. 16 Storage Scalability Containers Queue system Costs as First Class Metric
17. 17 Storage, storage, storage Price/GB/month Prices based on Google cloud 06/2022, Frankfurt (europe-west3) Standard Storage ($0.023) Nearline Storage ($0.013) Coldline Storage ($0.006) Archive Storage ($0.0025)
18. 18 Storage, storage, storage Standard Storage ($141,312) Nearline Storage ($79,872) Coldline Storage ($36,864) Archive Storage ($15,360) Example costs per year, 500 TB, $ Prices based on Google cloud 06/2022, Frankfurt (europe-west3)
19. Storage, storage, storage Example costs per year, 500 TB, $ 19 Standard Storage ($141,312) Nearline Storage ($79,872) Coldline Storage ($36,864) Archive Storage ($15,360) Prices based on Google cloud 06/2022, Frankfurt (europe-west3)
20. 20 Storage, storage, storage Price/GB/month Self hosted storage ($0.001/ redundant: $0.002) Prices based on Google cloud 06/2022, Frankfurt (europe-west3) Standard Storage ($0.023) Nearline Storage ($0.013) Coldline Storage ($0.006) Archive Storage ($0.0025)
21. Storage, storage, storage Example costs per year, 500 TB, $ Self hosted storage ($0.001/ redundant: $0.002) Self hosted storage ($6,758) 21 Standard Storage ($141,312) Nearline Storage ($79,872) Coldline Storage ($36,864) Archive Storage ($15,360) Prices based on Google cloud 06/2022, Frankfurt (europe-west3)
22. 22 Storage Scalability Containers Queue system Costs as First Class Metric
23. 23 Choosing the right queue system Message Broker Message Bus Apache Kafka - min. 4GB RAM (according to online discussions) - we know that smaller sizes are possible RabbitMQ - min. 256MB RAM (according to RabbitMQ) - it works best if the queues are empty Think of your requirements/use case: Lifetime (Persistence) Ordering Priority Messages Protocol (MQTT, AMQP, …) Routing Topics
24. 24 Storage Scalability Containers Queue system Costs as First Class Metric
25. 25 Optimized scaling of cloud costs Cost-aware scalability Right sizing of services Picking the right instance (more than 1.7 million combinations in AWS) Optimized servers for database, computing, graphics, storage, … Cost dimensions Network Data transfer Data storage IO Processing/Runtime Scale an event only after all dimensions have been calculated Tier requirements (Processing, event, data) Monitoring Realtime cost analysis Planning Enforcement of scaling actions
26. 26 Storage Scalability Containers Queue system Costs as First Class Metric
27. 27 Reduce cloud costs by using containers Just containers Simplifies coordination and cooperation between teams Build them on powerful machines Deploy them on small machines Use the smallest base containers (Alpine Image has 5MB) One-time configuration (vs. configuring servers/VMs) Low maintenance costs Base containers are open source and free to use Use multi-stage build to reduce container size Kill non-used environments with policies
28. 28 Storage Scalability Containers Queue system Costs as First Class Metric
29. 29 Costs as First Class Metric Storage Scalability Containers Queue system Costs as First Class Metric
30. 30 Our approach in numbers By applying all these steps we power our realtime AI end-to-end data pipeline per customer with: Storage Servers (20-40TB) 30-50 EUR per month DB Servers 10-15 EUR per month App Servers 5-10 EUR per month Gitlab Server 30 EUR per month AI Hardware (2-4x Nvidia 3060 RTX) connected via queue server paid once + electricity Price: 75-105 EUR per month
31. 31 Contact Brantner Digital Solutions GmbH Dr.-Franz-Wilhelm-Straße 2a 3500 Krems an der Donau also in Vienna ;) T: +43 664 145 67 54 E: rene.heinzl@brantner.com © BRANTNER ÖSTERREICH GMBH, 2022 We are looking forward to meeting you