Workshops Archive

Heterogeneous Energy Characterization of GPU-Powered Datalog Engines


Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices

Authors: Ahmedur Rahman Shovon (Argonne National Laboratory), Yihao Sun (Utah State University), Zhiling Lan (University of Illinois Chicago), Swann Perarnau (Argonne National Laboratory), Thomas Gilray (Washington State University), Kristopher Micinski (Syracuse University), and Michael E. Papka and Sidharth Kumar (University of Illinois Chicago)

Abstract: Datalog’s declarative rules support recursive analytics in databases, program analysis, and machine learning, yet the energy behavior of GPU-powered Datalog engines remains largely unexplored. We present the first heterogeneous CPU–GPU energy characterization of these systems, evaluating five engines across hash-based, index-nested-loop, sorted/hash-indexed, and DataFrame-based designs on NVIDIA A100 GPUs. We measure total CPU–GPU energy, hardware-counter-based instructions per joule, and energy–delay product using Powerlog, a lightweight harness that samples GPU power through NVML and CPU-package power through RAPL while running unmodified engines. Results show that the fastest configuration is not always the most energy-efficient. Engine choice changes total energy by up to 3×, while CPUs account for 32–55% of total energy. Scaling from one to four GPUs reduces runtime by up to 2.9× but increases energy by 24–42%. Our reproducible methodology and artifacts support energy-aware Datalog system design and fair cross-engine comparison.


Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing

Back to Full Workshop Archive Listing