Performance Evaluation and Workload Recharacterization of SRE Extended Kernel via Graph500 Benchmark

This technical report presents a comprehensive performance evaluation and workload recharacterization of the self-developed **Status-Relational Entropy (SRE) Extended Kernel under the Graph500 v2.1 benchmark specification. Deployed on a low-power, thermal-constrained mobile edge computing architecture (Intel Core i5-8265U, 16 GB DDR4 RAM, 1 MPI process coupled with 8 OpenMP threads), the benchmark evaluates a graph topology of SCALE = 12 ($4,096$ unique vertices, $65,536$ edges, and $134,752$ non-zero sparse matrix elements) across 64 independent test runs (NBFS = 64). Key performance telemetry reveals an optimized harmonic mean throughput of $724,534$ TEPS (sre_harmonic_mean_TEPS) and a peak throughput of $761,312$ TEPS (sre_max_TEPS), with a mean execution time of 0.185984 s and an average per-vertex processing latency of 45.4063. Through sparse data layout restructuring and OpenMP thread scheduling, the updated implementation eliminates long-tail latency outliers, reducing the maximum execution time from 0.391 s to 0.228 s and shrinking runtime standard deviation (σₜᵢₘₑ) by ~75%. Crucially, the report recharacterizes the SRE Kernel from traditional memory-bound Graph500 Breadth-First Search (BFS) to a compute-bound, global graph-matrix concurrent accelerator engine. Unlike localized BFS, the SRE Kernel enforces a $100%$ non-speculative traversal over all non-zero elements ($nnz = 134,752$) while concurrently executing high-order algebraic transformations—including localized degree calculations, matrix traces, Frobenius norms, and sub-space Krylov spectral sieving. This on-the-fly topological-algebraic fusion eliminates secondary memory sweeps, offering critical computational efficiency for Graph Neural Network (GNN) execution layers, real-time enterprise fraud detection, and lock-free relational simulation engines. This dataset and technical report present a comprehensive performance evaluation and workload recharacterization of the self-developed Status-Relational Entropy (SRE) Extended Kernel under the Graph500 specification. Deployed on an edge architecture (Intel Core i5-8265U, 16 GB RAM, 1 MPI process x 8 OpenMP threads), the benchmarking evaluates both baseline scale (SCALE=12) and large-scale parallel scalability up to SCALE=20 (1,048,576 vertices, 34,600,692 non-zero matrix elements). Empirical results at SCALE=20 demonstrate a breakthrough performance of 20,362,900 TEPS (sre_harmonic_mean_TEPS) with a mean execution time of 1.6992 s and a per-vertex processing latency of 1.62 μs. Demonstrating exceptional parallel efficiency, a 256x expansion in graph size (from SCALE=12 to 20) resulted in only a ~9.1x increase in execution time, driving a 28x boost in TEPS throughput while maintaining strict latency stability (stddev = 0.0328 s). Per-operator profiling indicates that sub-space spectral sieving (Op6, Krylov Lanczos iteration) accounts for 90.5% of total compute time. The accompanying result_20.txt log provides raw empirical telemetry verifying the deterministic edge traversal (stddev_nedge = 0) and high-density compute-bound graph-matrix acceleration.