Domain Project

We are happy to announce that our paper “A Bootstrapping Technique for Reducing the Costs of Machine Learning Models for Predicting Execution Times in IaaS Clouds” has been accepted at the 2025 ACM Symposium on Cloud Computing (SoCC’25)).

The study addresses a fundamental challenge in Infrastructure-as-a-Service (IaaS) environments: the high cost and time required to collect sufficient empirical data for training accurate Machine Learning (ML) models that predict application performance across diverse Virtual Machine (VM) configurations. To mitigate these overheads, the authors propose a bootstrapping technique that leverages the Universal Scalability Law (USL) to generate synthetic training samples. By integrating analytical modeling with ML, the approach enables the creation of robust performance predictors even when only a minimal number of real-world measurements are available.

The experimental validation, conducted on Amazon EC2 using a variety of benchmarks and 32 different VM types, demonstrates that this bootstrapping method significantly improves prediction accuracy under data-scarcity conditions. Within the context of the DOMAIN project, this work provides a scalable framework for optimizing resource provisioning and task scheduling. By reducing the reliance on extensive experimental campaigns, the proposed methodology facilitates the deployment of more agile and cost-effective cloud management strategies, aligning with the project’s broader goals of fostering innovation in distributed systems and performance engineering.