Standard Task Graph (STG) Multi-Solver Benchmark Results for Workflow Scheduling in Heterogeneous High Performance Computing (HPC) Systems - Phase I: Small Scale Edge Device

This dataset contains benchmark results from running six workflow scheduling solvers on Standard Task Graph (STG) workflow instances mapped onto simulatedhomogeneous (3-node) and heterogeneous (8-node) HPC system configurations. The dataset covers three graph sizes: rnc50 (50 tasks), rnc100 (100 tasks), and rnc300 (300 tasks), each in both homo and hetero modes (180 workflow instances per cell). Solvers included: MILP via PuLP (CBC), MILP via Gurobi, CP-SAT via OR-Tools, HEFT heuristic, and two GNNRL model variants (self and teacher). All runs are Small Scale Benchmark Tests in Edge Device (Phase I), executed on an Intel Core i5-1145G7 edge device (4 cores / 8 threads, 2.60 GHz base / 4.40 GHz boost, 15 W TDP, 16 GB RAM, Ubuntu 22.04.5 LTS, Python 3.10.12). Cluster-based validation on real HPC infrastructure (Phase II) will be published as a separate companion dataset. The primary dataset is in benchmark_solver_results_main.zip, which unpacks to main_results/ with one subfolder per solver. It contains 6,480 JSON files (180 instances per solver per cell, best result selected where multiple runs existed). Each JSON file records the objective value, makespan, resource utilisation, and per-task schedule for one workflow instance and one solver run. Analysis figures and summary statistics are in benchmark_figures.zip (21 PDF + 21 PNG figures + summary_stats.csv). Figures use the Okabe-Ito (2008) colorblind-safe palette with distinct hatches and markers for mono-print use. All solvers minimise the same weighted objective: Objective = alpha * makespan + beta * usage_term (alpha = beta = 1.0) Input workflow graphs are available from the companion STG dataset at DOI: 10.5281/zenodo.18927122. Solver scripts are available at: https://github.com/AasishKumarSharma/grapheonrl-benchmark These results serve as baseline reference points for PhD thesis research on Graph Neural Network Reinforcement Learning (GNNRL) for workflow scheduling at the University of Gottingen / GWDG.