Slides
As cloud computing is gaining more popularity, it is a critical problem for cloud users to select optimal configurations for their applications, to improve performance at the same budget or save cost while achieving required performance. From a measurement on 29 compute intensiveapplications, we observed performance diversity on Amazon EC2 instances. Instance size and processor type are key factors leading to such diverse performance. Intuitive rules of thumb are not applicable here. It is not always true that instances with larger sizes or newer processor types achieve better performance than instances with smaller sizes or older processors.
We proposed Nearest Neighbor shortlisting to select optimal configuration for a target application effectively and cost-efficiently. We select applications of similar characteristics with the target application and shortlist top configuration candidates based on these applications' performance; with some testing of the target application on these top configuration candidates, the optimal configuration is shortlisted. Our evaluation showed that Nearest Neighbor shortlisting is able to get the right configuration with low error and low cost.