WINTER: Adaptive Serverless Prewarming with Limited Invocation History — Reproducibility Artifact

This record contains the companion research data, software, and manuscript for WINTER: Adaptive Serverless Prewarming with Limited Invocation History by Guntak Kim, Jaehong Jung, and Gu-In Kwon (Department of Computer Science and Engineering, Inha University, Republic of Korea). The included manuscript is prepared for submission to IEEE Access. The DOI identifies the reproducibility artifact. WINTER combines a source-trained temporal representation, prototype initialization, and a locally updated ridge adapter for serverless prewarming. WINTER-G selects among prototype, adapted, and exponentially weighted moving-average predictions using observed age and invocation count. The archived evidence covers initial observation, continuous operation, workload shifts, and active-pool measurements using Azure Functions 2019/2021 and Huawei Cloud 2023 traces. Contents and scope. Artifact version 2026.09.28 retains the experimental snapshot of 2026.09.17 and the IEEE Access manuscript source revision ieee-access-20260922. It includes all 19 final case configurations and fixed action sets; function-level outcomes averaged over seeds and seed-level totals; source aggregate/contract records; model checkpoints and source partition identities; the reported controller-selection and representation-control evidence; live request/response and calibration measurements; and table/figure inputs. Provider-distributed datasets and the complete processed trace universes are excluded. The exact selected final replay input arrays, totaling approximately 16.4 MB, are retained to identify and replay the reported experiments. Large intermediate per-minute function/seed ledgers and feature, rate, and forecast caches are omitted. The supplied code supports regeneration within the documented replay scope; full training requires the upstream datasets and the additional setup described in the repository. Download and restore. The data payload totals 613,725,130 bytes (approximately 614 MB) in eight logical TAR.GZ volumes. Volumes 01 and 08 are supplied whole; volumes 02 through 07 are supplied as 65 binary parts of at most 8,000,000 bytes each. Download all 67 data files together with restore-data.py and data-bundle.json, and run python3 restore-data.py. The helper checks the transport files and original archive SHA-256 hashes and reconstructs the six split volumes; it does not extract them. Extract all eight complete TAR.GZ volumes and winter-code-ieee-access-20260928.zip into the same parent directory to restore winter-serverless/. Each complete TAR.GZ volume is extracted separately. README.md, README-DATA.md, data-bundle.json, and SHA256SUMS.txt provide the file list, restoration commands, scope, and integrity information. Validation. Recorded checks cover 16 code tests; input/action/contract verification for 19 cases; 3,666 paired contrasts and 53,921 numerical checks; 20 regenerated table fragments; 10 matching figure pages; four representative function-level fixed-policy replay checks; and independent verification of controller selection and representation-control results. This validation is not a full simulation-campaign rerun, model retraining, an end-to-end public-raw-data-to-results rebuild, or a new physical-cluster timing study. Exact procedures and limits are documented in VALIDATION_REPORT.md and release-validation.json. Manuscript. The standalone winter-manuscript-ieee-access-20260926.zip, main.pdf, and supplementary.pdf provide the complete source package, the supplied 13-page main text, and the supplied 18-page supplement. The validation report records the pagination difference produced by the older TeX Live 2019 build environment. The manuscript retains its acknowledgment of AI assistance in writing and LaTeX formatting. Licensing and attribution. The MIT License applies to the author-written software within the scope of LICENSE. CC BY 4.0 applies to author-generated measurements, model parameters, tables, and figures within the scope of LICENSE-DATA. Selected trace derivatives retain the upstream CC BY 4.0 attribution requirements. These licenses apply to their respective components; they do not grant a choice of license for every file. The manuscript and third-party template, font, logo, and dependency assets retain their applicable rights and notices. See LICENSE, LICENSE-DATA, THIRD_PARTY.md, and docs/DATA.md. Upstream dataset repositories are linked in Related works. Funding. This research was supported by the Ministry of Science and ICT (MSIT), Korea, under the National Program for Excellence in Software, supervised by the Institute of Information & Communications Technology Planning & Evaluation (IITP) in 2024 (Grant No. 2022-0-01127). This work was also supported by INHA University. Contact. Corresponding author: Gu-In Kwon, [e-mail].