Spot instance Strategies - AWS techies

6,176 views

Published on

Published in: Technology

Spot instance Strategies - AWS techies

  1. 1. Let’s discuss Spot Instances Spot Strategies
  2. 2. WHO AM I? Aater Suleman Part-time UT Professor Geek, Architect, Developer, Ops, DevOps … Co-founder & CEO Flux7 Labs ○ DevOps Solutions ■ ■ Deployments Cost/performance optimized large scale website (Ruby on rails, node.js, Django) and Hadoop deployments
  3. 3. Spot Instances Consumer chooses a maximum bid AWS charges market price @ start of the hour Can get terminated at any time
  4. 4. Spot Strategy Checklist Spot terminations Spot outages Launch failures Optimized price = min(spot price, on-demand price) Ability to auto-scale
  5. 5. Similarities with Auto-scaling Stateless instances Apps can tolerate momentary dip in capacity Apps can tolerate instance “swapping” Launch failures
  6. 6. Differences from auto-scaling? No programmable cooldown period
  7. 7. AppNeta’s (http://www.appneta.com/blog/aws-spot-instances/) Two ASGs (spot and main) SNS notifications on Spot ASGs Outside daemon: ○ Spot instance termination ⇒ on-demand instance
  8. 8. Analysis Spot terminations Spot outages Launch failures Optimized price Ability to auto-scale
  9. 9. Sanket’s Algorithm Two ASGs (spot and on-demand) ● Setup thresholds as follows: ASG: On-demand ASG: Spot Scale up: CPU > 75% Scale up: CPU > 60% Scale down: CPU < 55% Scale down: CPU < 45%
  10. 10. Analysis Spot terminations Spot outages Launch failures Optimized price Ability to auto-scale
  11. 11. Flux7 v1.0 Our own auto-scaling solution that allows hybrid spot and on-demand Step Up: If spot price < on-demand, launch spot Spot Launch failure: Launch on-demand Step Down: If on-demand is available, terminate it End of hour: Treat like Step Up
  12. 12. Analysis Spot terminations Spot outages Launch failures Optimized price Ability to auto-scale (in-house ASG)
  13. 13. Thank You

×