Work

Pre-registered design: energy and cost per query of CPU and GPU dataframe engines (Polars, Spark, cudf-polars, RAPIDS) on A100, H100, GH200 and B200 hosts and a CPU-only bare-metal host (Paper 3), locked 2026-11-16

Year: 2026
Type: article
Abstract: Pre-registration (locked design) for Paper 3. This record timestamps the locked design file docs/PAPER3_ENERGY_DESIGN.md (git tag paper3-design-lock in the repository below) and the pre-lock probe evidence it cites. No study query had been run when the design was locked; the probes used synthetic loads only (a CUDA spin kernel, nvidia-smi pollers and pinned CPU spin loops). Aim. Rigorous, reproducible evaluation methods and tools that help U.S. organizations choose cost- and energy-efficient CPU or GPU engines for analytical data processing. The study measures energy and cost per query of Polars and Apache Spark, each on the CPU and on the GPU, with scope-tagged energy (NVIDIA NVML energy counter on Lambda GPU hosts; Intel RAPL on a GCP c3-highmem-192-metal CPU-only host), a 10 s window rule with per-host measurement resolution fixed by pre-lock probes, integrity and telemetry gates, a Williams Latin-square order seeded by the lock date, and pre-specified break-even rules. It contributes a measurement protocol and break-even scales, not a vendor ranking. Files. PAPER3_ENERGY_DESIGN.md (the locked design); paper3-prelock-evidence.tar.gz (docs/paper3-evidence/, docs/PAPER3_INFRA.md and paper3_hosts.json at the lock commit); SHA256SUMS. Funding and conflict of interest. Lambda provided $10,000 and AWS $6,000 in compute credits for Paper 3; neither sponsor had any role in design, analysis, interpretation, review or the decision to publish. Use of AI tools. AI coding and writing assistants (Anthropic's Claude via Claude Code; OpenAI Codex in the earlier exploratory work) were used to write and test code, run and monitor the cloud experiments, draft the design and analysis plan, and draft and edit the text. The author directed the study, reviewed and approved the design, the analyses and the text, and takes full responsibility for the content. Workload notice. The TPC-H-derived workload is derived from the TPC-H Benchmark and as such is not comparable to published TPC-H Benchmark results, as the TPC-H-derived workload results do not comply with the TPC-H Benchmark Specification. The energy figures are not TPC-Energy metrics.
Institution: Boston University
Language: English
Open Access status: green
OpenAlex ID: W7234009885
FWCI:
Cites: 4
Cited by:
Related to: 10