caoyujia/adapdomain-edge-scheduling

AdapDomain Evaluation Supplement

This dataset accompanies the manuscript AdapDomain: Residual-Shape Scheduling for Multi-Tenant Edge Inference: An Ascend Server Study with Large Language Model Prefill and Decode.

Contents

Current live LLaMA-7B measurements (2026-10-16)

live_llama7b_20261016/ contains the summary measurements used for the current manuscript's LLaMA-7B tables and figure: stage profiles, phase-conditioned co-location slowdowns, multi-trace goodput with 95% confidence intervals, and burst-trace latency and SLO-violation results. These were measured on four Ascend 910B3 cards with LLaMA-7B FP16 using vLLM-Ascend. Use this dated folder for the current reported LLM results. The older 2026-10-10 supplementary archive below is retained as a historical snapshot and is not the source for these updated values.

Supplementary_AdapDomain_20261010.zip contains the manuscript's software, configurations, experiment outputs, and hardware/profile materials described in the archive. The project_dataset/ directory is an additional, separately organized dataset bundle uploaded from the AdapNPU project:

• project_dataset/azure_trace/: Azure Functions 2019 invocation trace files (invocations_d01.csv through invocations_d14.csv) and a compact NPZ cache.
• project_dataset/synthetic_traces/: six workload trace JSON files (uniform, bursty, diurnal, spike_random, cv_heavy, and nlp_heavy).
• project_dataset/profiling_data/benchmark_baseline.json and project_dataset/test_cv_batch1.npy: included unchanged from the supplied project dataset directory. Their provenance and relationship to particular manuscript results are not specified here.

The additional files are provided as a separate project-data bundle; their presence does not imply that every file was used to produce a reported manuscript result. Consult the manuscript and the accompanying code/configuration before interpreting them.

Trace provenance and attribution

The Azure invocation traces are from the Azure Functions Dataset 2019, distributed by the Azure Public Dataset project. The dataset description specifies the CC-BY Attribution License. Please retain attribution and cite:

Mohammad Shahrad, Rodrigo Fonseca, Inigo Goiri, Gohar Chaudhry, Paul Batum, Jason Cooke, Eduardo Laureano, Colby Tresness, Mark Russinovich, and Ricardo Bianchini. “Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider.” In Proceedings of the 2020 USENIX Annual Technical Conference (USENIX ATC 20), USENIX Association, Boston, MA, July 2020. See the official Azure Functions Dataset 2019 description for the canonical attribution and license details.

The synthetic trace files are project-generated workloads, as described in the project asset manifest. No independent license or provenance statement is asserted here for benchmark_baseline.json or test_cv_batch1.npy.

Reproduction

The implementation is available at github.com/JackyYujia/AdapDomain. The archive includes the software and configuration snapshot associated with the supplementary materials. Hardware-specific experiments require a compatible Ascend environment.

Citation

Please cite the accompanying AdapDomain manuscript and this dataset record. Dataset record: caoyujia (2026), AdapDomain Evaluation Supplement, Hugging Face Datasets, https://huggingface.co/datasets/caoyujia/adapdomain-edge-scheduling. For reproducibility, cite an immutable dataset revision rather than the moving main branch.

License

No single license is declared for the combined repository because it contains materials with different or incompletely documented provenance. The Azure trace files have the attribution terms stated above; check the relevant source and file-level notices before reusing other materials. Public visibility does not itself grant reuse rights.

Archive integrity

SHA-256 of Supplementary_AdapDomain_20261010.zip:

9C30018730C25B29034B8CA206E96865B651755D987789C5DD8B0BE7B0FDC7A1