6 intelligent-placement-of-datacenters
- 1. Intelligent placement of datacenters for Internet Services Inigo Goiriyz, Kien Lez, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presenter: Zafar Gilani
- 2. Introduction • Selection of suitable datacenter locations is very important. • Why? – Running and maintenance costs. – Network latency. – Environmental factors (renewable energy vs carbon-intensive).
- 3. Important considerations for location selection • Proximity to – population centers, – power plants, and – network backbones. • Source of electricity in the region. • Electricity, land and water prices. • Average temperatures of the location.
- 4. Framework
- 5. Framework for placement • Goal: – Minimize overall cost, while respecting response time, consistency and availability. • Objectives: – Formalize the process as a non-linear cost optimization problem. – Automated datacenter location selection process.
- 6. Framework: Parameters • Capital costs: investments made upfront. Type of capital cost Description Independent of Electricity, external networking. number of servers Maximum number of Land acquisition, datacenter construction, servers power delivery, backup, cooling systems. Actual number of Purchase of servers, internal networking. servers
- 7. Framework: Parameters • Operational costs: incurred during operation. Type of operational Description cost Actual number of Maintenance of equipment, external servers bandwidth usage. Utilization of hosted Electricity and water costs. servers
- 8. Framework: Parameters • Response time. • Consistency delay. • Availability. • CO2 emissions.
- 9. Framework: Optimization problem Placement of a Maximum number of Number of servers that datacenter at service population center c locaton d, either servers at location d. at location d. 1 or 0.
- 10. Framework: Optimization problem Placement of a Maximum number of Number of servers that datacenter at service population center c locaton d, either servers at location d. at location d. 1 or 0.
- 11. Framework: Solution approaches • Make it linear. Remove Sd and Pd,c. Use linear version of PBd,c is use of servers at location d to CAP_max. serve population center c, either 1 or 0. This is actual number of servers at each location d.
- 12. Framework: Solution approaches • Using Heuristics: 1. Use simple linear program to generate M1 datacenter networks for 1 to D datacenters. We have M1 * D configurations. 2. Use SBd (placement) and PBd,c (use to meet demand) to derive pre-set linear program. 3. Select most popular locations and run brute force.
- 13. Framework: Solution approaches • Simulated Annealing: – For each candidate solution we have values for each location d and population center c. – Optimization starts with a configuration and datacenter at each location. – Each iteration evaluates a neighboring configuration. – Iterate until no more cost reductions observed for n iterations.
- 14. Input data and datacenter characteristics for placement tool
- 15. Input data SE BI SL NY LA AU OR
- 16. Input data
- 17. Input data
- 18. Datacenter characteristics • Datacenter size, cooling and PUEs. – 8% power delivery losses. • Connection costs. – $500K/mile for transmission. – $480K/mile for fiber optic. – $1 per Mbps. 1Mbps per server. • Building costs. – As a function of maximum power: $15 per watt (small), $12 per watt (large). – Availability: 99.827%
- 19. Datacenter characteristics • Land cost. – 6K sq. ft. per MW • Water cost. – 24K gallons of water per MW per day. • Server and internal networking hardware. – $2K per server. – $20K per switch. • Staff costs. – An admin can manage 1K servers for an average salary of $100K/year.
- 20. Results from the tool, a few characterizations
- 22. Location characteristics: observations City PUE/Temp Land/Water Network CO2 cost cost emissions Austin H L L L Bismarck L L H H Los Angeles H H L L New York H H L L Orlando H H L L Seattle L H L L St. Louis H L H H
- 23. A case study: placing a datacenter network
- 24. Evaluation
- 25. Evaluating solution approaches Heuristic was run for 3 days and then forcefully terminated, results were extrapolated. OSA+LP1 is: •2x faster than Heuristic. •5x faster than Brute.
- 26. Datacenter placement tradeoffs: Latency 2x difference in price between desired latency of 33ms and 50ms $7.8M/month for latency 70ms or more
- 27. Datacenter placement tradeoffs: Cheaper to have 3 Tier Availability Overall Tier II II than 2 Tier IV datacenters are the datacenters. best option.
- 28. Datacenter placement tradeoffs: Consistency delay Consistency delay and latency are conflicting goals. Acceptable ranges for consistency delay and latency.
- 29. Datacenter placement tradeoffs: Green datacenters A network of 8 datacenters with 60K servers produces 8K tons of CO2/month. Will cost a lot more for lower latencies. With relatively higher latency of 70ms, it will cost $100K/month more for green energy.
- 30. Datacenter placement tradeoffs: chiller-less datacenters Avoiding chillers can reduce costs by 8% for latencies > 70ms.
- 31. Conclusion
- 32. In a nutshell • Intelligent placement of datacenters can save millions of $/€ . • Cost of networks of datacenters doubles when maximum acceptable response time is reduce from 50ms to 35ms.
- 33. Intelligent placement of datacenters for Internet Services Inigo Goiriyz, Kien Lez, Jordi Guitart, Jordi Torres, and Ricardo Bianchini Presenter: Zafar Gilani