• Very related to AdaRes
  • It seems that they need to be able to change the VM resource allocation in a way that doesn’t hurt service quality?
  • Figure 1: Why does the confidence interval exploded after 8 tests?
  • Bayesian optimizer seems more reasonable than RL? Need to know more about Bayesian optimizer.
  • How to set the optimization goal (>0.5 happiness index)? Should this be a parameter set by customer or cloud provider?
  • How to read Figure 2?
  • I think Serverless might have more potential on automatic tuning. (easier to control resource allocation; more frequent feedback)
  • This paper has better evaluation than AdaRes. E.g. Image recognition using TensorFlow model; Video processing