HPC-Driven Relational Database Operators with Advanced GPU Memory Management

This project demonstrates GPU-accelerated relational database operators (Join and Group-By) using CUDA C++. It features:

• Parallel kernels optimized for coalesced memory access
• Use of pinned memory for efficient host-to-device transfers
• Performance verification comparing CPU and GPU implementations
• Scalability tests across varying dataset sizes
• A modular design for future extensions

Features

• Parallel Kernels: Implemented in joinKernel and sumByKeyKernel to ensure each thread accesses contiguous data.
• Profiling & Optimization: Structured to reduce warp divergence and memory contention.
• Optimized Memory Layout: Utilizes pinned memory (cudaMallocHost) for faster data transfers.
• Performance Verification: CPU and GPU operations are compared using C++ <chrono> for timing.
• Scalability Tests: Command-line parameters allow testing with different dataset sizes.

Requirements

• NVIDIA CUDA Toolkit
• A compatible NVIDIA GPU

Running the Project

./HPC_RelOps [numRowsA] [numRowsB] [maxKeyRange] [repeats]

Example:

./HPC_RelOps 100000 100000 50000 3

Project Structure

HPC_Relational_Ops/
├── README.md         # Project documentation
├── Makefile          # Build instructions
└── src/
    └── HPC_RelOps.cu # Complete C++/CUDA source code for the project