EfiMon: A Process Analyser for Granular Power Consumption Prediction

TL;DR

The paper addresses the challenge of granular energy accounting in HPC for AI workloads by introducing EfiMon, a non-invasive, agnostic monitor that collects per-process instruction histograms, CPU/RAM usage, and system power via adapters. It proposes a regression-based energy model that maps PSU power $P_{ ext{PSU},t}$ to instruction distributions $h_k^{p}$ and CPU usage $w_p$ with a static intercept, enabling non-isolated predictions for a Process of Interest (PoI). Experimental results on dual-socket AMD and Intel nodes show robust per-process power estimation with low relative error (approximately 2.2%–4.4%), validating applicability in shared-resource HPC environments. This work enables fine-grained energy accounting and energy-aware optimization in HPC and cloud-based acceleration services without requiring process isolation.

Abstract

High-performance computing (HPC) and supercomputing are critical in Artificial Intelligence (AI) research, development, and deployment. The extensive use of supercomputers for training complex AI models, which can take from days to months, raises significant concerns about energy consumption and carbon emissions. Traditional methods for estimating the energy consumption of HPC workloads rely on metering reports from computing nodes power supply units, assuming exclusive use of the entire node. This assumption is increasingly untenable with the advent of next-generation supercomputers that share resources to accelerate workloads, as seen in initiatives like Acceleration as a Service (XaaS) and cloud computing. This paper introduces EfiMon, an agnostic and non-invasive tool designed to extract detailed information about process execution, including instructions executed within specific time windows and CPU and RAM usage. Additionally, it captures comprehensive system metrics, such as power consumption reported by CPU sockets and PSUs. This data enables the development of prediction models to estimate the energy consumption of individual processes without requiring isolation. Using a regression-based mathematical model, our tool is able to estimate single processes' power consumption in isolated and shared resource environments. In shared scenarios, the model demonstrates robust performance, deviating by a maximum of 2.2% on Intel-based machines and 4.4% on AMD systems compared to non-shared cases. This significant accuracy showcases EfiMon's potential for enhancing energy accounting in supercomputing, contributing to more efficient and energy-aware optimisation strategies in HPC.

Figures (5)

• Figure 1: Efimon software stack. The architecture uses the interface-adapter pattern to decouple the dependencies and keep the API uniform and agnostic.
• Figure 2: Mean probability of the instruction types for each experiment measured using Linux Perf through EfiMon.
• Figure 3: Total CPU and PSU consumption of all experiments at different CPU usage from 0 to 100% (multiple threads). The total power comprises the two CPU sockets.
• Figure 4: Model prediction performances of different power consumption estimation models contrasted against the true value from IPMI (PSU) measurements. All models consider the CPU usage logarithmic.
• Figure 5: Assessment of the prediction in the absence and presence of noise for different levels of parallelism (number of cores).