Parallel Sparse Matrix Solver

This project implements a parallel sparse matrix solver using CUDA, featuring different optimization strategies including Level Analysis (LA) and Cache-Aware (CA) approaches.

Prerequisites

• CUDA Toolkit (version 12.8 or compatible)
• NVIDIA GPU with CUDA support
• GCC/G++ compiler
• Make
• NVIDIA Nsight Systems (nsys) for performance profiling (optional)

Project Structure

.
├── include/         # Header files
├── src/            # Source files
├── data/           # Test matrices
├── build/          # Build artifacts
├── Makefile        # Build configuration
└── run.sh          # Execution script

Building the Project

1. Make sure you have CUDA toolkit installed. The default path is set to /opt/nvidia/hpc_sdk/Linux_x86_64/2023/cuda/12.8. If your CUDA installation is in a different location, modify the CUDA_HOME variable in the Makefile.

2. Build the project:

make clean    # Clean previous build artifacts
make         # Build the project

This will create an executable named run_sparse in the root directory.

Running the Project

The project comes with a convenient run.sh script that runs the solver on multiple test matrices with different optimization strategies.

Available Test Matrices

• onetone1
• onetone2
• bcircuit
• G2_circuit
• hcircuit
• parabolic_fem

Running Options

1. Run with default settings:

./run.sh

2. Run with a specific size parameter:

./run.sh <size>

The script will:

• Run each matrix 20 times
• Use the LEVEL optimization strategy
• Generate performance metrics for analysis

Manual Execution

You can also run the solver manually:

./run_sparse <matrix_file> [optimization_strategy] [size]

Where:

• <matrix_file>: Path to the matrix file (e.g., "data/onetone1/onetone1.mtx")
• [optimization_strategy]: Optional optimization strategy (e.g., "LA_OPT")
• [size]: Task granularity size

Performance Profiling with Nsight Systems

To get detailed performance metrics using NVIDIA Nsight Systems:

nsys profile --stats=true ./run_sparse <matrix_file> [optimization_strategy] [size]

For example:

nsys profile --stats=true ./run_sparse ./data/onetone1/onetone1 LA_OPT

Results

The solver compares its results with serial execution for validation. The console output shows the running time and verification status:

Input: ./data/onetone1/onetone1.mtx
Algorithm: LA_OPT
Granularity: 2048
[CUDA LA_OPT Solve Time] 114.96 ms
[PASS] GPU == CPU

When running with nsys, the solver generates performance metrics and analysis files:

• .nsys-rep files: NVIDIA Nsight Systems performance reports
• .sqlite files: Performance data in SQLite format

Cleaning Up

To clean build artifacts:

make clean

References

Helal, Ahmed E., et al. "Adaptive task aggregation for high-performance sparse solvers on GPUs." 2019 28th International Conference on Parallel Architectures and Compilation Techniques (PACT). IEEE, 2019.