RA-SpMM Benchmark Graphs: 26 Real + 25 Synthetic Sparse Matrices for GNN-SpMM Evaluation

This Zenodo record archives the 51-graph evaluation suite used in the FGCS paper "Regime-Aware Sparse Matrix Multiplication for Graph Neural Network Workloads on GPUs" (Afridi & Lee, 2026). The suite combines: 26 real-world graphs covering six structural categories (Sparse Uniform, Sparse Skewed, Dense Small, Dense Large-Scale, Community, Mixed/Irregular), drawn from SNAP, the Open Graph Benchmark (OGB), and PyTorch Geometric.25 procedurally-generated synthetic graphs (5 per regime × 5 regimes) that stress under-represented regions of the (M, d, CVd) feature space.</ul> The archive is structured to drop directly into the RA-SpMM GitHub repository. After cloning the code repo, downloading this archive and running tar -xzf ra_spmm_data_v1.tar.gz --strip-components=1 at the repo root populates datasets/ and fgcs_results/synthetic/ where the included paper_datasets.json, paper_combined_datasets.json, and paper_synthetic_datasets.json manifests resolve them. Bundle SHA-256:eedcdc6285ce33a3af4e18ea8bd14d73cb43c2582c9ae362ee6bdc980f0a604f Code repository: https://github.com/tariqaf/RA-SpMMCitation: please cite the FGCS paper (when published). Per-graph attribution to upstream sources (SNAP, OGB, PyG) is provided in the included ATTRIBUTIONS.md file inside the archive.