Carbon optimisation of AI training workload in cloud environment

Sammanfattning

Artificial Intelligence (AI) has grown exponentially in recent years due to the development of deep learning and large language models (LLMs), which have increased computational requirements and energy consumption. AI models require a large amount of computing resources, long training times, and GPU clusters, causing an increase in carbon emissions and other environmental impacts. As AI continues to grow, it is time to consider more sustainable, environmentally friendly approaches to AI models. Training AI models requires the use of data centres, which consume significant amounts of energy, materials, and water. This can cause serious environmental issues and impact people. However, there are other data centres that are more eco-friendly or have lower carbon emissions, depending on factors such as wind and sunlight. The problem is that current cloud scheduling algorithms primarily focus on performance and cost, often neglecting carbon emissions and environmental impact. This thesis explores the idea of an eco-responsible, look-ahead multi-objective scheduling framework explicitly designed for delay-tolerant AI training workloads in order to reduce environmental impact while maintaining acceptable cost levels. For this research, we will use CloudSim, a cloud computing simulation framework that enables us to experiment without requiring much data or real-world cloud infrastructure. In this research, we are going to compare traditional scheduling algorithms, such as First-Fit, with a proposed eco-responsible scheduling approach that incorporates carbon emissions, cost, and performance as optimisation metrics. The expected outcome is to achieve a well-balanced result in terms of carbon emissions and performance while keeping an acceptable cost. Results show that the carbonaware look-ahead policy significantly outperforms the baseline schedulers, managing to lower carbon emissions while keeping acceptable costs. Additionally, adding dynamic runtime with a low amount of reallocation adds a marginal carbon and cost reduction of approximately 2%. This happens because the initial placement is highly effective at positioning higher workloads in green datacenters, while migrating the less heavy workloads. This proves that cloud infrastructures can minimise carbon emissions without compromising costs.