Extending the Life of HPC Systems in Resource Constrained Environments: Mapping Productivity-Energy Trade-offs in Memory-Bound Workloads via DVFS and Core Scaling on Repurposed Hardware
Jul 2026· Practice and Experience in Advanced Research Computing· 0 citations· 18 references
Computer Science
TL;DR
This work examines the impact of limiting active cores on repurposed nodes and introduces deep C-state power-gating, fully saturated workloads, and hardware-level power measurements to address viability in complex applications such as OpenFOAM.
Abstract
Many resource-constrained environments rely on repurposed hardware for High-Performance Computing, shifting costs from capital expenditure to operational energy costs. As a result, evaluating system viability in resource-constrained environments warrants a shift from measuring raw performance to evaluating Productivity-to-Energy efficiency. Memory-bound workloads are particularly impacted by the memory wall, where stalled processors waste energy. While Dynamic Voltage and Frequency Scaling is widely used, systematic core scaling remains largely overlooked. This work examines the impact of limiting active cores on repurposed nodes. Presenting Phase 1 preliminary results, initial HPCG benchmarking demonstrates that targeted core deactivation yields a 62.5% improvement in PTE efficiency over maximum-performance baselines. To address viability in complex applications such as OpenFOAM, a second phase introduces deep C-state power-gating, fully saturated workloads, and hardware-level power measurements. The resulting framework provides a practical, software-driven approach to lowering Total Cost of Ownership and advancing sustainable High-Performance Computing in resource-constrained settings.
MEPOWER is proposed, a flexible, model-based approach to exposing compute/data movement imbalance that characterizes the fine-grained memory behavior of parallel workloads that demonstrates a reduction in EDP on a range of HPC benchmarks with minimal impact on execution time when compared to the standard OS/hardware-managed power control mechanism.
Nanda Velugoti, Joseph Manzano, Andrés Márquez et al.· 0 citations
Distributed HPC and LLM workloads increasingly require efficient communication for scalability, yet growing data movement has become a major performance bottleneck. Communication compression can reduce this overhead and complement execution-level optimizations, but its benefits remain difficult to assess because existing benchmarks lack support for diverse backends, realistic datasets, application-specific accuracy metrics, and overlap-induced resource contention. We present CC-Bench, a lightweight, extensible, and application-oriented benchmark suite for evaluating communication compression under realistic execution conditions. CC-Bench uses declarative application-environment modeling to decouple profiling logic from communication libraries, datasets, and fidelity metrics, enabling portable cross-library evaluation. It further combines function-level interception and hardware counter monitoring to characterize per-phase latency, hardware utilization, numerical fidelity, and computation interference. With representative datasets from HPC and LLM workloads, CC-Bench evaluates three compression-enabled communication libraries on CPU and GPU clusters, revealing accuracy-performance trade-offs and bottlenecks to guide practical deployment and optimization.
This work presents a data-driven framework that leverages historical job traces to estimate the impact of resource modifications on queue performance, and introduces the Weighted Wait-Time Score (WWS), a bounded metric that captures both typical and tail wait-time behavior.
Bipin Gaikwad, Shraddha Singh, M. Joshi et al.· Practice and Experience in A...· 0 citations
Energy efficiency is a critical concern in modern heterogeneous HPC systems, where CPUs, GPUs, and dense cooling infrastructure increase both idle and peak energy demand. This paper presents a power-centric observability framework deployed on the NSF REPACSS system that enables non-intrusive and consistent energy analysis across facility, rack, node, and job scopes. The framework collects instantaneous power telemetry exclusively from out-of-band hardware sources, including in-row cooling units, rack power distribution units, and compute nodes accessed via iDRAC. These measurements are stored in a time series database and processed through a power-centric energy model that aligns heterogeneous telemetry in time and space. A scalable query engine and web API provide a uniform interface for deriving energy metrics across infrastructure and job contexts, while a lightweight Slurm epilog integration enables optional job level energy reporting without application modification or runtime instrumentation. Using this unified telemetry and query framework, we conduct a multi-scope energy characterization of the NSF REPACSS data center. The results show stable facility efficiency, efficient in-row cooling behavior, distinct differences in idle and peak power between CPU and GPU nodes, and diverse non-CPU energy contributions in representative workloads. Together, these findings demonstrate that unified out-of-band telemetry combined with a power-centric energy model, enables practical cross-scope energy observability in production HPC systems.
Yongjia Zhao, Jie Li, Chenxu Niu et al.· Practice and Experience in A...· 1 citation
Three fundamental design principles are revealed that provide design-space guidance for architects designing the next generation of memory-accelerated LLM systems.
Corey Lammie, Hadjer Benmeziane, W. Simon et al.· 0 citations
