Optimizing Amazon SageMaker Workloads with Predictive Compute Type Selection Strategies

Effective resource provisioning is a critical aspect of optimizing cloud infrastructure, ensuring efficient resource allocation, and managing costs judiciously. One key element in this process is predicting the most suitable "Compute Type" for various workloads, as different types exhibit varying efficiencies in handling specific tasks. The precision in selecting the appropriate compute type not only leads to enhanced performance but also contributes to cost savings. AWS offers a broad range of compute types designed to satisfy various requirements about cost, scalability, and performance. Because these instances are suitable for specific use cases and workloads, choosing the right compute type is essential. Effective compute type selection is essential for managing costs and allocating resources as efficiently as possible. Precise forecasts lead to enhanced functionality, financial savings, and increased energy efficiency. In this paper, we have four kinds of compute types available in AWS - Accelerated Compute Types, Compute Optimized Instances, Memory Optimized Instances, and Standard Instances. We have proposed an automated approach for prediction of most appropriate compute types for specific workloads, reducing the manual configuration. In order to create the predictive model, we have used the Support Vector Machines (SVM) classifier and the ensemble methods. The SVM model balances F1-scores across classes and obtains an amazing overall accuracy of 84.49%. Conversely, AdaBoost performs worse, ranking 69.10% accurate and having comparatively lower F1-scores, particularly for "Memory Optimized Instances." With a remarkable accuracy of 99.42% and high F1-scores across all classes, XGBoost is particularly good at classifying 'Compute Type' using the features that are provided.