An Analytical and Extended Cost-Effective Resource Provisioning Framework in IaaS Clouds under Online Learning
Preprint
Abstract
Cloud vendors such as Amazon EC2 offer three types of purchase options: reserved, on-demand and spot instances. A important problem for all users is determining the way of utilizing all pricing options to minimize the cost of processing all jobs while respecting the response-time targets. The application of online learning to this scenario is interesting in that it imposes no restriction of a priori statistical knowledge of workload and spot prices and achieves a good performance close to that of the best policy of the used set. So far, only spot and on-demand instances can be addressed by this approach and we enable it to address all pricing options and self-owned instances, which brings to users the opportunity to further reduce their cost. Importantly, we also lay some mathematical foundation for taking a holistic view to analyze and design a set of cost-optimal/effective policies that determine how many self-owned, spot and on-demand instances are assigned to each job and the final cost of completing all jobs. Finally, simulations are done, showing a markedly cost reductions when different combinations of self-owned, reserved, spot, and on-demand instances are considered.Keywords
All Related Versions
This publication has 5 references indexed in Scilit:
- IaaS reserved contract procurement optimisation with load predictionFuture Generation Computer Systems, 2015
- Cost-Minimizing Online VM Purchasing for Application Service Providers with Arbitrary Demands2015
- Optimal Online Multi-Instance Acquisition in IaaS CloudsIEEE Transactions on Parallel and Distributed Systems, 2014
- Optimal Bids for Spot VMs in a Cloud for Deadline Constrained Jobs2012
- Optimization of Resource Provisioning Cost in Cloud ComputingIEEE Transactions on Services Computing, 2011