GraphZero

High-Performance, Zero-Copy Graph Engine for Massive Datasets on Consumer Hardware.

GraphZero is a C++ graph processing engine with lightweight Python bindings designed to solve the "Memory Wall" in Graph Neural Networks (GNNs). It allows you to load and sample 100 Million+ node graphs (like ogbn-papers100M) and their massive feature matrices on a standard 16GB RAM laptop—something standard libraries like PyTorch Geometric (PyG) or DGL cannot do.

The Problem

GNN datasets can be massive. ogbn-papers100M contains 111 Million nodes, 1.6 Billion edges, and gigabytes of node embeddings.

• Standard approach (PyG/NetworkX): Tries to load the entire graph structure and all node features into RAM before training begins.
• The Result: MemoryError (OOM) on consumer hardware. You need 64GB+ RAM servers just to load the data.

The Solution:

GraphZero abandons the "Load-to-RAM" model. Instead, it uses a custom Zero-Copy Architecture:

• Memory Mapping (mmap): The graph and its features stay on disk. The OS only loads the specific "hot" pages needed for computation into RAM via page faults.
• Compressed CSR (.gl): A custom binary format that compresses raw edges by ~60% (30GB CSV $\to$ 13GB Binary).
• Columnar Tensor Store (.gd): A raw, C-contiguous binary format for node features that instantly translates to PyTorch tensors without memory allocation.
• Parallel Sampling: OpenMP-accelerated random walks that saturate NVMe SSD throughput, using thread-local RNGs to eliminate lock contention.

🏆 Benchmarks: GraphZero vs. PyTorch Geometric

Task: Load ogbn-papers100M (56GB Raw) and perform random walks. Hardware: Windows Laptop (16GB RAM, NVMe SSD).

Metric	GraphZero (v0.2)	PyTorch Geometric
Load Time	0.000000 s ⚡	FAILED (Crash) ❌
Peak RAM Usage	~5.1 GB (OS Cache)	>24.1 GB (Required)
Throughput	1,264,000 steps/s	N/A
Status	✅ Success	❌ OOM Error

Proof of Performance

Left: GraphZero loading instantly and utilizing OS Page Cache. Right: PyG crashing with Unable to allocate 24.1 GiB.

📦 Installation

GraphZero is available on PyPI:

pip install graphzero

🚀 Quick Start

1. Convert Your Data (Topology & Features)

GraphZero uses high-efficiency binary formats. Convert your generic CSV lists once.

example edges.csv, weights are optional:

src,dst,weight
0,1,0.5
1,2,1.0

import graphzero as gz

# 1. Convert Topology (Edges & Weights) to .gl
gz.convert_csv_to_gl(
    input_csv="dataset/edges.csv", 
    output_bin="graph.gl", 
    directed=True
)

# 2. Convert Node Features to .gd (Float32, Int64, etc.)
gz.convert_csv_to_gd(
    csv_path="dataset/features.csv",
    out_path="features.gd",
    dtype=gz.DataType.FLOAT32
)

2. High-Speed Sampling & Zero-Copy Tensors

Once converted, the graph and its multi-gigabyte feature matrix are instantly accessible without consuming RAM.

import graphzero as gz
import numpy as np

# TOPOLOGY
# Zero-Copy Load (Instant)
g = gz.Graph("graph.gl")

# Define Start Nodes
start_nodes = np.random.randint(0, g.num_nodes, 1000).astype(np.uint64)

# Parallel Biased Random Walk (Node2Vec style: p=1.0, q=0.5)
walks = g.batch_random_walk(
    start_nodes=start_nodes, 
    walk_length=10,
    p=1.0, 
    q=0.5
)

# FEATURES
# Zero-Copy Feature Load (Instant)
fs = gz.FeatureStore("features.gd")

# Get a perfect 2D Numpy/PyTorch Tensor mapping directly to the SSD
# RAM used: 0 Bytes!
node_features = fs.get_tensor() 

print(f"Graph loaded. Feature Matrix Shape: {node_features.shape}")

⚙️ Under the Hood

GraphZero is built for Systems & GNN enthusiasts.

• Core: C++20 with nanobind for Python bindings.
• Parallelism: Uses #pragma omp with thread-local deterministic RNGs.
• IO: Direct CreateFileMapping (Windows) and mmap (Linux) calls with alignment optimization (4KB/2MB pages).

🌟 Current Features List (v0.2)

GraphZero currently supports the following high-performance ML capabilities:

Graph Structural Engine

• Instant Ingestion: Fast mmap-backed loading of directed, undirected, and weighted graphs.
• Zero-Copy CSR: Custom .gl binary format for dense, continuous memory alignment and 64-byte CPU cache line optimization.
• Thread-Safe Sampling: OpenMP-accelerated batch_random_walk_uniform and batch_random_fanout.
• Biased Walks (Node2Vec): Hardware-optimized Alias Table generation for $O(1)$ weighted sampling (batch_random_walk with p and q parameters).
• Fault-Tolerant: Automatic handling of dead-ends (sinks) and out-of-bounds nodes.

Graph Data Engine

• Columnar Tensor Store: Custom .gd binary format for storing $N \times F$ feature matrices.
• Strong Typing: Native C++ template dispatching supporting FLOAT32, FLOAT64, INT32, and INT64.
• Zero-Copy Bridge: Direct translation of mmap pointers to Numpy/PyTorch multidimensional arrays.

🗺️ Roadmap

• v0.3 (The Algorithmic Core): High-performance analytics engine adding OpenMP-accelerated Parallel BFS/DFS, PageRank, and Connected Components.

• v0.4 (Dynamic Updates): Breaking the immutable CSR barrier via an LSM-Tree/Adjacency List memory overlay to allow real-time edge/node insertions.

• v0.5 (Production Hardening): ACID-compliant safety for multi-process PyTorch training using Reader-Writer Locks, Write-Ahead Logging (WAL), and graceful exception handling.

📄 License

MIT License. Created by Krish Singaria (IIT Mandi).