Efficient Join Algorithms For Large Database Tables in a Multi-GPU Environment.

Area of Science:

• Computer Science
• Database Systems
• Parallel Computing

Background:

• Relational join processing is fundamental to database management systems.
• Graphics Processing Units (GPUs) show promise for accelerating relational joins.
• Existing GPU join algorithms struggle with very large datasets and multi-GPU environments.

Purpose of the Study:

• To address the challenge of handling large input data in relational join processing using multiple GPUs.
• To explore the benefits of multi-GPU environments for join processing.
• To design efficient multi-GPU join algorithms overcoming CPU-GPU data transfer limitations.

Main Methods:

• Proposed three distinctive multi-GPU join algorithms: nested loop, global sort-merge, and hybrid joins.
• Focused on optimizing data transfer between CPUs and GPUs for large table joins.
• Conducted extensive experiments on multiple databases and hardware configurations.

Main Results:

• Demonstrated high scalability of the proposed algorithms with increasing data size.
• Achieved significant performance boosts through the utilization of multiple GPUs.
• Outperformed existing join algorithms, showing speedups of up to 25X over multi-core CPUs and 2.8X over single GPUs.

Conclusions:

• The developed multi-GPU join algorithms effectively handle large datasets.
• Multi-GPU systems offer substantial performance advantages for relational join processing.
• These algorithms represent a significant advancement in database acceleration on modern hardware.