Adaptive Cost Efficient Framework for Cloud-Based Machine Learning

Machine learning is an increasingly important form of cognitive computing, making progress in several application areas. Machine learning often involves big data sets and is computationally challenging, requiring efficient use of resources. The use of cloud computing as the platform for machine learning offers advantages of scalability and efficient use of hardware. It may, however, be difficult to provision appropriate cost-effective resources for a machine learning task. Our experiments have shown that there can be radical differences between different datasets and different algorithms on the same dataset. The cloud-based machine learning framework presented here aims to provide multiple levels of efficient use of resources and uses a high-level cost model to deal with overall cost-efficiency with respect to cloud service providers. The cost model allows evaluation of trade-offs and supports the choice of appropriate provider resources based on user-defined criteria. A user may choose to prioritize performance, prioritize cost or specify a cost-performance balance. An Amazon AWS cost model for instances is used to illustrate the practical benefits of using the approach - it is seen that large savings can be made by employing this job-specific monitoring and cost-performance analysis. The method can provide all the information for a comparison across different cloud service providers as well as comparisons across the Amazon AWS offerings.