Hybrid CPU-GPU Sorting Mechanisms for Large-Scale Out-of-Memory Data

open_access
Open Access
Author:
Choi, Leah
Area of Honors:
Computer Engineering
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Robert Collins, Thesis Supervisor
Anand Sivasubramaniam, Thesis Honors Advisor
Keywords:
CUDA
Hybrid sort
Sorting
GPU
CPU
UVM
Abstract:
In the modern technology landscape, efficient data sorting is crucial for the performance of numerous applications in fields ranging from machine learning to large-scale database management. As datasets continue to grow, traditional processing methods often struggle to keep pace, necessitating the development of more sophisticated technologies to handle these volumes efficiently. This thesis investigates the comparative performance of GPU-heavy versus hybrid CPU-GPU sorting methods, pivotal for addressing the challenges posed by large datasets. Our findings reveal that while a GPU-centric approach provides high throughput for smaller datasets less than 25GB, it faces significant performance decline when processing larger datasets due to memory constraints and increased data transfer overheads. A critical memory threshold, around 25GB, is identified, beyond which GPU performance deteriorates sharply due to excessive page faults and data transfer overheads associated with Unified Virtual Memory (UVM). Conversely, a balanced hybrid approach, utilizing a 50/50 CPU to GPU load factor, offers a strategic solution by distributing workloads to avoid GPU memory saturation and minimizing data movement. This method not only leverages the strengths of both the CPU and GPU but also addresses their limitations, enhancing overall system performance and stability across varying data sizes. The research underscores the importance of strategic workload distribution and memory management in large-scale data environments, providing a sustainable solution to the complexities of sorting large datasets.