Performance Optimization in the Cloud for Spot and Burstable Resources

  • Date September 16, 2015
  • Hour 11.30 am
  • Room GSSI Room D
  • Speaker Dr. Daniel Dubois (Imperial College, London, UK)
  • Area Computer Science

In recent years some of the main cloud providers are offering new pricing models for renting their resources. The most popular ones are the spot model, in which resources are offered to the highest bidder; and the burstable model, in which resources have a variable speed that is based on their past utilization. Spot and burstable resources are typically cheaper than traditional ones with constant hourly price and constant performance, but there is a risk to lose a spot resource when its price is overbid by other users, and to lose up to 90% of performance when heavily using a burstable resource for a long time.

This talk discusses how to predict the performance of an application that uses these types of cloud resources and how to use these predictions to make decisions on: (1) which and how many resources to rent; (2) how to map the application components to the rented resources; and (3) what spot price bids to use in order to minimize the total bid price while maintaining an acceptable level of performance.