Banking on decoupling

Abstract

Cloud providers are auctioning their excess capacity using dynamically priced virtual instances. These spot instances provide significant savings compared to on-demand or fixed price instances. The users willing to use these resources are asked to provide a maximum bid price per hour, and the cloud provider runs the instances as long as the market price is below the user's bid price. By using such resources, the users are exposed explicitly to failures, and need to adapt their applications to provide some level of fault tolerance. In this paper, we expose the effect of bidding in the case of virtual HPC clusters composed of spot instances. We describe the interesting effect of uniform versus non-uniform bidding in terms of both the failure rate and the failure model. We propose an initial attempt to deal with the problem of predicting the runtime of a parallel application under various bidding strategies and various system parameters. We describe the relationship between bidding strategies and programming models, and we build a preliminary optimization model that uses real price traces from Amazon Web Services as inputs, as well as instrumented values related to the processing and network capacities of cluster instances on the EC2 services. Our results show preliminary insights into the relationship between non-uniform bidding and application scaling strategies.