حاتم ابوالقاسمي

In recent years, deep learning, as a subset of machine learning, has gained widespread attention due to its ability to perform complex tasks such as image recognition, natural language processing, an‎d behavior analysis. However, deploying these models in container-based environments faces challenges including scalability, resource management, an‎d workload adaptability. In particular, the continuous deman‎d for heavy computational resources an‎d the dynamic, variable nature of workloads highlight the necessity for intelligent scheduling solutions. This thesis presents a novel workload-aware scheduling framework for containerized deep learning environments. Utilizing a soft actor-critic reinforcement learning algorithm, the framework dynamically makes resource allocation decisions by simultaneously analyzing system metrics such as CPU, GPU, an‎d memory usage along with workload-specific features like model complexity an‎d training progress. Additionally, a complementary component is designed to optimize the input parameters of each job—including batch size, number of training epochs, an‎d requested resource amounts—prior to scheduling. To eva‎luate the system’s performance, comprehensive simulations were conducted using real trace data from the Alibaba cluster over a homogeneous 13-node infrastructure. Experimental results demonstrate that the proposed framework reduces the average job completion time by up to 38%, improves training accuracy to 85%, an‎d decreases the average training error to 0.27. Moreover, CPU consumption was reduced by 20%, GPU consumption by 18%, an‎d memory usage by 15%, while resource distribution became more balanced. These results indicate the high efficiency of the proposed approach in enhancing the productivity an‎d scalability of deep learning workloads in modern container-based infrastructures.

يادگيري عميق , گنجانه سازي , مقياس پذيري , چالش هاي استقرار

deep learning , containerization , scalability , deployment challenges

Hatam Abolghasemi

Mehrdad Ashtiani