Improving Power-Measurement Reliability Through Cross-Platform Calibration and Comparison of Software APIs

Abstract

Accurate power measurement is fundamental to systems research, compute optimization, and financial planning, even small errors can scale into substantial operational and monetary losses in large-scale data-center environments. Although software-based power measurement Application Programming Interfaces (APIs) are widely adopted for their convenience and ease of integration, our study shows that the APIs display consistent deviations from hardware-level ground-truth measurements. These inaccuracies are not merely minor discrepancies; they can meaningfully distort empirical conclusions and, in some cases, invalidate comparative analysis in both academic research and industry reporting.

In this thesis, we systematically quantify the unreliability of software-based approaches and present a practical and generalizable method to correct their measurement errors. Our correction procedure is evaluated on multiple devices, operating systems, and GPU architectures, demonstrating substantial robustness and portability. Using only accessible and inexpensive measurement equipment, the proposed approach reduces API error and significantly improves the stability of recorded power-consumption trends. We further validate its effectiveness across several widely used software APIs to ensure methodological consistency.

Beyond documenting the hidden cost of relying on uncorrected software power data, this work also provides an adaptable solution that researchers and practitioners can incorporate into their evaluation pipelines. Improves the reliability of power-consumption metrics and contributes to more reproducible and cost-aware system-level analysis across diverse computing platforms.