Advanced Techniques for Distributing and Timing Artificial Intelligence Based Heavy Tasks in Cloud Ecosystems
Year: 2024
Type: article
Abstract: Today, the proliferation of Artificial Intelligence (AI) workloads has brought to light a requirement for nuanced cloud infrastructures that can efficiently process AI-heavy workloads. The results of this study, focusing on workload allocation and scheduling in cloud systems that are experiencing the additional challenge emanating from AI-intensive jobs. In this work, we aimed to investigate the prevailing techniques and limitations in current methodologies, with an eye toward designing new methods that can improve how AI resources are allocated and scheduled in cloud environments. The reason for this problem is that AI-related workloads are of diverse resources (why), change due to their nature, and require scalability. Artificial intelligence is also extremely computationally intensive and has a diversification of compute needs making distributing resources efficiently another beast. In addition, the nature of these activities is dynamic, which means responsive approaches are needed to cope with changing computing demands over time. In addition, given that both AI models and datasets are growing in complexity as well size this makes scalability an imperative for cloud environments to be able operate effectively. In this literature review, we will describe traditional and state-of-the-art task allocation systems to give a complete picture of their pros and cons. Moreover, we discuss the scheduling strategy used for managing tasks requiring more significant AI and provide an extensive review of state-of-the-art. To address these challenges, we introduce a sophisticated framework that emphasizes the deployability of resources and scheduling by leveraging machine learning tools, as well as more efficient job transfer methods. The system is designed to allocate resources elastically and match the dynamic requirements of AI based workloads. This is accomplished by employing machine learning algorithms to predict workload information and introducing task migration mechanisms for adjusting changes in workloads. The study concludes through an empirical assessment of the proposed solutions in a virtual environment using different datasets. We objectively measure the performance of our methods with respect to traditional approaches using important metrics such as throughput, latency and resource utilization. Such analysis can provide useful insights on cost effective deployment of AI rich workloads onto cloud infrastructures. It helps current initiatives to improve the scalability and performance of cloud environments as AI applications multiply.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4400596168