Skip to content

GPU Processing

The InSAR Norway Change Detection toolkit includes legacy GPU-accelerated processing capabilities for users with compatible NVIDIA hardware.

Overview

GPU processing can significantly speed up change detection analysis for large datasets. The GPU-accelerated code is available in the gpu_legacy/ directory.

Legacy Code

The GPU processing scripts are legacy code and are no longer actively maintained. They are provided for users with existing GPU setups and specific performance requirements.

Prerequisites

Hardware Requirements

  • NVIDIA GPU with CUDA support (Volta™ or higher)
  • CUDA Compute Capability 7.0 or higher
  • Minimum 8GB GPU memory (16GB+ recommended for large datasets)
  • Linux distributions with glibc>=2.28

Software Requirements

  • Python 3.12
  • NVIDIA CUDA Toolkit
  • RAPIDS libraries (cuDF, CuPy)
  • Conda package manager

Setup

1. Check GPU Compatibility

Bash
1
2
3
4
5
# Check if NVIDIA GPU is available
nvidia-smi

# Verify CUDA installation
nvcc --version

2. Create GPU Environment

Bash
1
2
3
4
5
6
# Navigate to GPU legacy directory
cd gpu_legacy/

# Create GPU-specific environment
conda env create -f environment.yml
conda activate insar-env  # Name specified in gpu_legacy/environment.yml

3. Verify GPU Setup

Bash
1
2
3
4
5
6
# Check if NVIDIA GPU is available
nvidia-smi

# Test GPU availability in Python
python -c "import cupy; print('CuPy GPU support available')"
python -c "import cudf; print('RAPIDS cuDF available')"

Available GPU Scripts

hybrid_detector.py (GPU Version)

CPU/GPU hybrid change detection processing with automatic optimization.

Bash
1
2
cd gpu_legacy/
python hybrid_detector.py --path /data/insar --maximum --cpu-threads 16 --gpu-threshold 30000

Key Parameters: - --maximum: Use maximum available resources
- --cpu-threads: Number of CPU threads for small files (default: 6) - --gpu-threads: Number of GPU threads for large files (default: 2) - --gpu-threshold: Point threshold for GPU processing (default: 10000) - --chunk-size: GPU batch size (default: 100000) - --gpu-memory-fraction: Target GPU memory utilization (default: 0.9) - --coherence: Minimum coherence threshold (default: 0.7)

insar-change-detector.py

Legacy GPU-focused processing script.

Bash
1
python insar-change-detector.py --path /data/insar --help

insar-visualizer.py (GPU Version)

Visualization with potential GPU acceleration for large datasets.

Bash
1
python insar-visualizer.py --results change_detection_results.csv --num-points 5 --selection mixed

Performance Comparison

Dataset Size CPU Time GPU Time Speedup
Small (< 1GB) 5 min 2 min 2.5x
Medium (1-5GB) 30 min 8 min 3.7x
Large (5-20GB) 2 hours 25 min 4.8x
Very Large (20GB+) 8+ hours 1.5 hours 5.3x

GPU Memory Management

The GPU scripts use a chunking strategy to process large datasets efficiently. The --chunk-size parameter controls how many points are processed simultaneously.

Memory Guidelines

Chunk Size Recommendations: - 8GB GPU: 12,000-14,000 points - 16GB GPU: 20,000-22,000 points
- Other GPUs: Start with 10,000 and increase gradually

Memory Optimization

Bash
1
2
3
4
5
6
7
8
# Monitor GPU memory usage
nvidia-smi -l 1

# Adjust chunk size for your GPU
python hybrid_detector.py --path /data --maximum --chunk-size 12000

# For memory-constrained systems
python hybrid_detector.py --path /data --maximum --chunk-size 8000

Troubleshooting Memory Issues

CUDA Out of Memory: - Reduce --chunk-size parameter - Close other GPU applications - Monitor with nvidia-smi

Troubleshooting

Common GPU Issues

CUDA Out of Memory - Reduce batch size - Use gradient checkpointing - Clear GPU cache between runs

GPU Not Detected - Verify NVIDIA drivers are installed - Check CUDA toolkit installation - Ensure GPU is not used by other processes

Slow GPU Performance - Check for thermal throttling - Verify sufficient power supply - Monitor GPU utilization with nvidia-smi

Migration to CPU Version

If you're currently using GPU processing and want to migrate to the maintained CPU version:

  1. Data Compatibility: Results are compatible between versions
  2. Algorithm Mapping: GPU algorithms map to CPU equivalents
  3. Performance: Modern CPU version is well-optimized
  4. Maintenance: CPU version receives active updates and bug fixes

Migration Example

Bash
1
2
3
4
5
6
7
# Old GPU command
cd gpu_legacy/
python hybrid_detector.py --path /data --moderate --gpu

# New CPU equivalent
cd ../
python hybrid_detector-cpu.py --path /data --moderate --cpu-threads 8

Future Considerations

  • CPU Performance: Modern multi-core CPUs provide excellent performance
  • Maintenance: CPU version receives active development
  • Compatibility: Broader hardware compatibility
  • Simplicity: Easier setup and deployment

Support

For GPU-specific issues: 1. Check the gpu_legacy/README.md for detailed setup instructions 2. Verify hardware compatibility 3. Consider migrating to the actively maintained CPU version

Recommendation

For new projects, we recommend using the CPU-based processing pipeline which is actively maintained and provides excellent performance on modern hardware.