Restricted Research - Award List, Note/Discussion Page
Fiscal Year: 2023
1536 The University of Texas at Arlington (143424)
Principal Investigator: Mohammad Islam,mislam@uta.edu,(817) 272-3785
Total Amount of Contract, Award, or Gift (Annual before 2011): $ 400,000
Exceeds $250,000 (Is it flagged?): Yes
Start and End Dates: 9/1/22 - 8/31/25
Restricted Research: YES
Academic Discipline: Department of Computer Science & Engineering
Department, Center, School, or Institute: none
Title of Contract, Award, or Gift: Utilizing Conducted Electromagnetic Interference (EMI) for Low-Cost Server-Level Power Monitoring in Data Centers
Name of Granting or Contracting Agency/Entity:
National Science Foundation (NSF)
CFDA Link: NSF
47.041
Program Title:
Communications, Circuits, and Sensing-Systems (CCSS)
CFDA Linked: Engineering Grants
Note:
(SAM Category 1.1.1.) Overview: Data centers are notoriously energy-hungry and, therefore, mandate efficient management where power monitoring plays a critical role. However, the significant cost barrier due to dedicated sensors has been hindering power monitoring's penetration to the server level. In this project, we develop the first-of-its-kind low-cost server-level power monitoring in data centers that extracts server power usage information from the conducted electromagnetic interference (EMI) generated by server power supplies. We significantly lower the cost by eliminating the need for a dedicated power meter for each server. Instead, we use the voltage measurement from a single point using a single sensor to provide server-level power consumption. The enabling insight for our power metering approach is that the power factor correction (PFC) circuits in server power supply spill their power usage information into its power network through conducted EMI. Specifically, the PFC circuits mitigate the power line harmonics generated by the power supply. However, as a side effect, it creates high-frequency (20kHz to 150kHz) current ripples that can be extracted from the voltage measurement at the equipment that the current flows through (in our case the cluster PDU). More importantly, these ripples changes with the server power and can be used to estimate the server power consumption. Since the power signature is generated at a high-frequency range, it is not adversely affected by harmonics generated from the power grid. Further, these power signatures exhibit approximate orthogonality' as different servers independently generate their signatures at different frequencies and occupy a narrow band (120Hz). Our preliminary experiments with server power supplies from Dell, SuperMicro, and Lenovo show very promising results where we observe server-generated EMI which reveals the server power. Nonetheless, there are several technical challenges towards realizing a conducted EMI-based power monitoring system in data centers. Propagation of conducted EMI and its extraction from voltage measurements in a data center is not well studied. Moreover, the impact of data center architecture and environment on EMI is unexplored. On the other hand, the extraction of many servers' power from a single sensor poses fundamental source separation challenges. Not to mention, we also need to convert the EMI to Watt as well as map the EMI-based readings to physical servers. Meanwhile, our high-frequency EMI extraction also poses computation challenges and requires efficient sensor design. Intellectual Merits: In light of the aforementioned challenges, we list our contribution in this proposal as follows. We augment our experiment results on conducted EMI from computer servers with theoretical understanding. We characterize how conducted EMI propagates in data centers and why we can extract it using voltage measurement. We develop a novel approach of extrapolation-based training of a semi-supervised system to extract server-level power from a single sensor. Besides, we adopt an offline data-driven approach for EMI to Watts conversion and a software-based technique to efficiently map EMI-based power readings to the physical servers. We develop a computation-efficient sensor design utilizing the superheterodyne principle to significantly reduce the rate of the data stream. Using our software implementation, we also introduce the concept of virtual meters for servers and tunable temporal resolution.
Discussion: No discussion notes