PEN: balancing energy and performance in NVM-based systems via DVFS with minimal wear-out
Edge devices are highly resource-constrained, requiring careful management of both energy efficiency and latency. Dynamic Voltage-Frequency Scaling (DVFS) is widely used to balance this trade-off and improve system performance. However, when Non-Volatile Memory (NVM) serves as the main memory, offering high density and low power consumption, DVFS can accelerate memory wear and reduce the useful lifetime. To address this, we propose PEN : a p olicy for e nergy–performance –efficient operation of N VM-based systems via DVFS while respecting memory lifetime constraints. PEN minimizes refreshes caused by system deceleration and takes advantage of acceleration to enable faster write modes, enhancing both lifetime and performance beyond what DVFS alone can achieve. Unlike previous approaches, PEN accounts for multicore-induced non-deterministic slowdowns and incorporates a low-overhead fallback refresh mechanism to preserve data integrity. Our results show that PEN improves Energy Delay Product (EDP) and NVM lifetime by 24.5% and 33%, respectively, compared to the state of the art.
Authors
- Nils Hölscher (ORCID: https://orcid.org/0000-0002-6397-5134)
- Christian Hakert (ORCID: https://orcid.org/0000-0001-9992-9415)
- Lokesh Siddhu (ORCID: https://orcid.org/0000-0002-5312-8679)
- Hassan Nassar (ORCID: https://orcid.org/0000-0003-1566-8997)
- Lars Bauer (ORCID: https://orcid.org/0000-0003-0253-4594)
- Henkel Jorg
- Miran Tobar (ORCID: https://orcid.org/0009-0008-1203-6197)
- Jian-Jia Chen
- Paul Wedeck
Institutions
- Karlsruhe Institute of Technology (DE)
- Indian Institute of Technology Guwahati (IN)
- TU Dortmund University (DE)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1038/s41598-026-80419-y
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Deutsche Forschungsgemeinschaft