PPMT: Parallel Power Measurements Tool
Date Issued
June 10, 2026
Author(s)
Description
The PPMT library is available:
GitLab link: https://www.ict.inaf.it/gitlab/energy/ppmt/-/releases/v0.1.0
Software Version: v0.1.0
Publication Date: June 2026
Abstract
The Parallel Power Measurements Tool (PPMT) is a C++ library designed for real-time, hardware-aware energy profiling of HPC applications across CPU and GPU architectures, supporting both serial and MPI-distributed environments. Built on top of the PMT library, PPMT extends its capabilities to distributed-memory systems by introducing topology-aware MPI communication hierarchies (socket, node, world) and device-specific energy aggregation strategies: maximum energy per socket (for RAPL-based CPU measurements) and summed energy across GPUs (for NVML/rocm-smi counters). This enables accurate, scalable attribution of energy consumption to specific computational phases, such as data transfers, kernels, or idle periods, with a negligible impact on the application performance. PPMT delivers hierarchical reports (per-rank, socket, node, global) that reveal energy hotspots and imbalances in large-scale simulations, directly supporting optimization of energy-efficiency in astrophysical codes. This work is a direct implementation of the USC-C Computing strategy at INAF, which prioritizes Green Computing through systematic energy profiling, hardware-aware algorithm co-design, and sustainable HPC operations. By quantifying the energy cost of scientific computation per watt, PPMT empowers researchers to maximize science output while minimizing environmental and economic footprint, aligning with the global imperative for energy-efficient exascale computing.
Report number
393
Rights
open.access
File(s)![Thumbnail Image]()
Loading...
Name
PPMT.pdf
Type
Postprint
Size
2.39 MB
Format
Adobe PDF
Checksum (MD5)
5fd6ae6876f686bd748ad319b4acb84b