Data and Code for Adaptive Model Selection for Resource-Constrained Image Classification Services Under Concurrent Workloads

Data and reproducibility code supporting the manuscript "Adaptive Model Selection for Resource-Constrained Image Classification Services Under Concurrent Workloads" by Jinpeng Wang. Version 1.1.0 adds E020, a second-host x86/Linux replication on CPU and NVIDIA RTX 3090. The frozen selector was evaluated with two workers under Poisson and bursty arrivals at 0.50, 0.80, and 1.05 times device-specific fixed-ResNet capacity, using ten paired repetitions per condition (28,800 formal requests). Routing reduced mean end-to-end latency by 15.7%–43.8% across CPU conditions but increased it by 99.3%–447.8% across RTX 3090 conditions. GPU board energy per request was 1.09–2.09 times the fixed-ResNet value. All formal paired effects had the same device-specific direction (exact two-sided sign-flip p=0.001953 per condition); no E020 request exceeded two seconds. The package preserves original predictions, routing decisions, feature ablations, empirical-trace simulations, live-serving records, accepted and rejected runs, E020 per-request records and power samples, aggregate statistics, environment locks, protocols, provenance, checksums, and executable code organized by manuscript Sections 5.1–5.9. The controlled E020 run completed with return code 0; 609 output-manifest entries were verified by SHA-256 remotely and after controlled return. Derived data and documentation: CC BY 4.0. Source code: MIT. ImageNetV2 images and pretrained torchvision weights are not redistributed and must be obtained from their original sources. Zenodo stores this release as eight numbered ZIP parts because the available upload path repeatedly interrupted the 59.2 MB single-file transfer. Download part00 through part07 and follow README_REASSEMBLY.txt; the reconstructed ZIP is byte-identical to the validated release (SHA-256 3454360b6c204f1d4a1761ded5100d8785ebe4bb62af3ae8107c131c0c9d7adf; MD5 a4b6aba2532d26d7587de655a18c6c65). Contact: [e-mail].