Abstract
This study explores how GPU resources can be used more efficiently for AI workloads running on shared clusters. Many existing schedulers leave GPUs idle or create long wait times for jobs. To address this, I developed three scheduling approaches—Hybrid Priority, Predictive Backfill, and Smart Batch—and tested them in a simulated multi-node environment. Each scheduler focuses on a specific improvement: Hybrid Priority balances fairness and efficiency, Predictive Backfill fills upcoming resource gaps, and Smart Batch groups similar jobs to reduce overhead. I compared them with standard schedulers like FIFO and SJF using metrics such as utilization, throughput, and wait time. The results show that the new schedulers reduce idle GPU time and improve fairness and overall performance. This work provides a practical framework for testing GPU scheduling strategies and helps guide future designs for efficient resource management in AI systems.
Advisor
Guarnera, Drew
Department
Computer Science
Recommended Citation
Mamirov, Akhmadillo, "Optimizing GPU utilization for AI workloads" (2025). Senior Independent Study Theses. Paper 11756.
https://openworks.wooster.edu/independentstudy/11756
Keywords
GPU, AI
Publication Date
2025
Degree Granted
Bachelor of Arts
Document Type
Senior Independent Study Thesis
© Copyright 2025 Akhmadillo Mamirov