Designing Self-Tuning Split-Map-Merge Applications for High Cost-Efficiency in the Cloud

Cloud platforms are attractive for executing large concurrent applications that require access to a pool of resources for concurrently executing the partitions of their workloads. Historically, application designers have tuned concurrent applications for specific hardware and platforms. But such approaches are not viable in cloud platforms as applications can be deployed on a variety of platforms and the operating environments can vary in each deployment. In this work, we argue and demonstrate that concurrent applications in cloud platforms must be self-tuning. First, we show that applications must incorporate a model of the overheads of operation. Second, we show that applications must determine their resource requirements and tune their operation to the operating conditions using estimations from the model. We build two self-tuning applications, E-Sort and E-MAKER, and demonstrate their ability to achieve high cost-efficiency by determining the right scale of partitions and resources to use for operation and adapting their behavior according to the characteristics of the deployed environment.