RAPIDS Framework Overview¶

📖 RAPIDS Documentation - Complete RAPIDS ecosystem reference

RAPIDS Ecosystem¶

Core Libraries¶

cuDF - GPU DataFrames: - GPU-accelerated pandas-like DataFrame library - Data loading, filtering, groupby, joins, string operations - 10-100x faster than pandas for large datasets - Primary data manipulation tool in RAPIDS

cuML - GPU Machine Learning: - GPU-accelerated scikit-learn-like ML library - Classification, regression, clustering, dimensionality reduction - Compatible API with scikit-learn - Significant speedup for training and inference

cuGraph - GPU Graph Analytics: - GPU-accelerated graph analysis library - PageRank, BFS, community detection, similarity - NetworkX-compatible interface - Scales to billion-edge graphs

Dask-cuDF - Multi-GPU DataFrames: - Distributed cuDF across multiple GPUs - Scales beyond single-GPU memory limits - Integrates with Dask scheduler - Partition-based parallel processing

Supporting Libraries¶

cuSpatial: GPU-accelerated spatial and geospatial analysis cuCIM: GPU-accelerated image processing for computational pathology cuSignal: GPU-accelerated signal processing RMM: RAPIDS Memory Manager for GPU memory allocation cuPy: GPU-accelerated NumPy replacement

CPU to GPU Mapping¶

GPU-Accelerated Data Science Benefits¶

Performance¶

• Massively parallel processing on GPU cores
• High memory bandwidth (GPU HBM vs CPU DDR)
• Efficient for columnar data operations
• Batch processing of large datasets

Ease of Adoption¶

• API compatibility with popular CPU libraries
• Minimal code changes for migration
• Interoperability with pandas, NumPy, scikit-learn
• Docker containers with pre-configured environments

Scalability¶

• Single GPU for medium datasets
• Multi-GPU with Dask-cuDF for larger datasets
• Multi-node scaling with Dask distributed scheduler
• Integration with Spark for enterprise data pipelines

Installation and Setup¶

Conda Installation¶

conda create -n rapids python=3.10
conda activate rapids
conda install -c rapidsai -c conda-forge -c nvidia \
  rapids=24.02 cuda-version=12.0

Docker Container¶

# Pull RAPIDS container from NGC
docker pull nvcr.io/nvidia/rapidsai/base:24.02-cuda12.0-py3.10

# Run with GPU access
docker run --gpus all -it -p 8888:8888 \
  nvcr.io/nvidia/rapidsai/base:24.02-cuda12.0-py3.10

Requirements¶

• NVIDIA GPU (Pascal or newer, compute capability 6.0+)
• CUDA toolkit (version matching RAPIDS release)
• Sufficient GPU memory for datasets
• Linux operating system (Ubuntu, CentOS, RHEL)

📖 RAPIDS Installation Guide - Setup instructions

Memory Management with RMM¶

RAPIDS Memory Manager (RMM)¶

• Manages GPU memory allocation for all RAPIDS libraries
• Pool allocator reduces allocation overhead
• Managed memory for automatic CPU-GPU data movement
• Memory logging for debugging

Configuration¶

import rmm
from rmm.allocators.cupy import rmm_cupy_allocator

# Pool allocator (pre-allocate GPU memory)
rmm.reinitialize(pool_allocator=True, initial_pool_size=2**30)

# Managed memory (allows oversubscription)
rmm.reinitialize(managed_memory=True)

Memory Best Practices¶

• Use pool allocator for production workloads
• Monitor GPU memory with nvidia-smi or RMM logging
• Process data in chunks for datasets larger than GPU memory
• Release unused DataFrames to free GPU memory
• Use efficient data types (int32 vs int64 when possible)

Interoperability¶

pandas Integration¶

import pandas as pd
import cudf

# pandas to cuDF
gpu_df = cudf.from_pandas(pd_df)

# cuDF to pandas
pd_df = gpu_df.to_pandas()

NumPy/CuPy Integration¶

import cupy as cp
import cudf

# cuDF to CuPy array
arr = gpu_df['column'].values

# CuPy to cuDF
gpu_df = cudf.DataFrame({'col': cp.array([1, 2, 3])})

cuda_array_interface¶

• Standard protocol for sharing GPU memory between libraries
• Zero-copy data sharing between cuDF, cuML, cuPy
• Avoids unnecessary data copies on GPU

Key Exam Concepts¶

• RAPIDS component mapping to CPU equivalents
• Benefits of GPU-accelerated data science
• Installation methods (conda, Docker, pip)
• RMM configuration for memory management
• Interoperability between RAPIDS and pandas/NumPy
• cuda_array_interface for zero-copy sharing