skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Understanding GPU Power. A Survey of Profiling, Modeling, and Simulation Methods

Abstract

Modern graphics processing units (GPUs) have complex architectures that admit exceptional performance and energy efficiency for high throughput applications.Though GPUs consume large amounts of power, their use for high throughput applications facilitate state-of-the-art energy efficiency and performance. Consequently, continued development relies on understanding their power consumption. Our work is a survey of GPU power modeling and profiling methods with increased detail on noteworthy efforts. Moreover, as direct measurement of GPU power is necessary for model evaluation and parameter initiation, internal and external power sensors are discussed. Hardware counters, which are low-level tallies of hardware events, share strong correlation to power use and performance. Statistical correlation between power and performance counters has yielded worthwhile GPU power models, yet the complexity inherent to GPU architectures presents new hurdles for power modeling. Developments and challenges of counter-based GPU power modeling is discussed. Often building on the counter-based models, research efforts for GPU power simulation, which make power predictions from input code and hardware knowledge, provide opportunities for optimization in programming or architectural design. Noteworthy strides in power simulations for GPUs are included along with their performance or functional simulator counterparts when appropriate. Lastly, possible directions for future research are discussed.

Authors:
 [1];  [1];  [1]
  1. Oak Ridge National Lab. (ORNL), Oak Ridge, TN (United States)
Publication Date:
Research Org.:
Oak Ridge National Lab. (ORNL), Oak Ridge, TN (United States). Oak Ridge Leadership Computing Facility (OLCF)
Sponsoring Org.:
USDOE
OSTI Identifier:
1326472
Grant/Contract Number:
AC05-00OR22725
Resource Type:
Journal Article: Accepted Manuscript
Journal Name:
ACM Computing Surveys
Additional Journal Information:
Journal Volume: 49; Journal Issue: 3; Journal ID: ISSN 0360-0300
Publisher:
Association for Computing Machinery (ACM)
Country of Publication:
United States
Language:
English
Subject:
29 ENERGY PLANNING, POLICY, AND ECONOMY; 32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATION; gpu; power; energy; profiling; simulator

Citation Formats

Bridges, Robert A., Imam, Neena, and Mintz, Tiffany M. Understanding GPU Power. A Survey of Profiling, Modeling, and Simulation Methods. United States: N. p., 2016. Web. doi:10.1145/2962131.
Bridges, Robert A., Imam, Neena, & Mintz, Tiffany M. Understanding GPU Power. A Survey of Profiling, Modeling, and Simulation Methods. United States. doi:10.1145/2962131.
Bridges, Robert A., Imam, Neena, and Mintz, Tiffany M. 2016. "Understanding GPU Power. A Survey of Profiling, Modeling, and Simulation Methods". United States. doi:10.1145/2962131. https://www.osti.gov/servlets/purl/1326472.
@article{osti_1326472,
title = {Understanding GPU Power. A Survey of Profiling, Modeling, and Simulation Methods},
author = {Bridges, Robert A. and Imam, Neena and Mintz, Tiffany M.},
abstractNote = {Modern graphics processing units (GPUs) have complex architectures that admit exceptional performance and energy efficiency for high throughput applications.Though GPUs consume large amounts of power, their use for high throughput applications facilitate state-of-the-art energy efficiency and performance. Consequently, continued development relies on understanding their power consumption. Our work is a survey of GPU power modeling and profiling methods with increased detail on noteworthy efforts. Moreover, as direct measurement of GPU power is necessary for model evaluation and parameter initiation, internal and external power sensors are discussed. Hardware counters, which are low-level tallies of hardware events, share strong correlation to power use and performance. Statistical correlation between power and performance counters has yielded worthwhile GPU power models, yet the complexity inherent to GPU architectures presents new hurdles for power modeling. Developments and challenges of counter-based GPU power modeling is discussed. Often building on the counter-based models, research efforts for GPU power simulation, which make power predictions from input code and hardware knowledge, provide opportunities for optimization in programming or architectural design. Noteworthy strides in power simulations for GPUs are included along with their performance or functional simulator counterparts when appropriate. Lastly, possible directions for future research are discussed.},
doi = {10.1145/2962131},
journal = {ACM Computing Surveys},
number = 3,
volume = 49,
place = {United States},
year = 2016,
month = 9
}

Journal Article:
Free Publicly Available Full Text
Publisher's Version of Record

Citation Metrics:
Cited by: 1work
Citation information provided by
Web of Science

Save / Share:
  • Most wireless sensor network (comprising of thousands of WSNs) applications require operation over extended periods of time beginning with their deployment. Network lifetime is extremely critical for most applications and is one of the limiting factors for energy-constrained networks. Based on applications, there are wide ranges of different energy sources suitable for powering WSNs. A battery is traditionally used to power WSNs. The deployed WSN is required to last for long time. Due to finite amount of energy present in batteries, it is not feasible to replace batteries. Recently there has been a new surge in the area of energymore » harvesting were ambient energy in the environment can be utilized to prolong the lifetime of WSNs. Some of the sources of ambient energies are solar power, thermal gradient, human motion and body heat, vibrations, and ambient RF energy. The design and development of TEGs to power WSNs that would remain active for a long period of time requires comprehensive understanding of WSN operational. This motivates the research in modeling the lifetime, i.e., power consumption, of a WSN by taking into consideration various node and network level activities. A WSN must perform three essential tasks: sense events, perform quick local information processing of sensed events, and wirelessly exchange locally processed data with the base station or with other WSNs in the network. Each task has a power cost per unit tine and an additional cost when switching between tasks. There are number of other considerations that must also be taken into account when computing the power consumption associated with each task. The considerations includes: number of events occurring in a fixed active time period and the duration of each event, event-information processing time, total communication time, number of retransmission, etc. Additionally, at the network level the communication of information data packets between WSNs involves collisions, latency, and retransmission, which result in unanticipated power losses. This report focuses rigorous stochastic modeling of power demand for a schedule-driven WSN utilizing Institute of Electrical and Electronics Engineers 802.11 and 802.15.4 communication protocols. The model captures the generic operation of a schedule-driven WSN when an external event occurs, i.e., sensing, following by processing, and followed by communication. The report will present development of an expression to compute the expected energy consumption per operational cycle of a schedule-driven WSN by taking into consideration the node level activities, i.e., sensing and processing, and the network level activities, i.e., channel access, packet collision, retransmission attempts, and transmission of a data packet.« less
  • Stochastic approximation methods have found extensive and diversified applications. Recent emergence of networked systems and cyber-physical systems has generated renewed interest in advancing stochastic approximation into a general framework to support algorithm development for information processing and decisions in such systems. This paper presents a survey on some recent developments in stochastic approximation methods and their applications. Using connected vehicles in platoon formation and coordination as a platform, we highlight some traditional and new methodologies of stochastic approximation algorithms and explain how they can be used to capture essential features in networked systems. Distinct features of networked systems with randomlymore » switching topologies, dynamically evolving parameters, and unknown delays are presented, and control strategies are provided.« less
  • Here, the development of renewable biofuels is a global priority, but success will require novel technologies that greatly improve our understanding of microbial systems biology. An approach with great promise in enabling functional characterization of microbes is activity-based protein profiling (ABPP), which employs chemical probes to directly measure enzyme function in discrete enzyme classes in vivo and/or in vitro, thereby facilitating the rapid discovery of new biocatalysts and enabling much improved biofuel production platforms. We review general design strategies in ABPP, and highlight recent advances that are or could be pivotal to biofuels processes including applications of ABPP to cellulosicmore » bioethanol, biodiesel, and phototrophic production of hydrocarbons. We also examine the key challenges and opportunities of ABPP in renewable biofuels research. The integration of ABPP with molecular and systems biology approaches will shed new insight on the catalytic and regulatory mechanisms of functional enzymes and their synergistic effects in the field of biofuels production.« less
  • This two-part report presents a review of the modeling and analysis of harmonic propagation in electric power systems, along with practical considerations and sample case studies. Part 1 of the report concentrates on the theoretical aspects of harmonics modeling and simulation. Concepts and characteristics of power system harmonics, modeling of harmonic sources and network components, and techniques for network-wide harmonic analysis are discussed.
  • Previously published work on applied impact damping typically relates to SDOF models or simple MDOF models such as the classical cantilever beam. Structural models often require an extremely large number of DOF with mode shapes that are generally very complex. Dynamics simulation of these typically becomes both complicated and time consuming. The nonlinear behavior of impact dampers further complicates such simulation in that standard linear solutions are not possible. The primary objective in this research extends previous work by applying impact dampers to MDOF structures that are modeled with general three-dimensional beam finite elements. Modal-based models of the MDOF systemsmore » and efficient impact damper tracking algorithms were also developed that significantly reduced CPU time for simulation. Significant among the objectives was obtaining an impact damper design for the MDOF casing structure of the Space Shuttle Main Engine (SSME), High-Pressure Oxygen Turbo-Pump (HPOTP), subject to pump rotor shaft unbalance. Impact damper performance is based on suppression of vibration at casing critical frequencies for rotor speed ranges, at rotor full speed, and very high unbalance to simulate a defect such as losing an impeller blade fragment or a cracked bearing. Simulations show significant reductions in vibration at the casing critical frequencies and very high unbalance levels while little or no improvement was observed off resonance. Additionally, the previous work with an experimental rotor bearing system (RBS) and impact damper was modeled using the developed modal-based methods. Simulation of the resulting model response shows remarkable agreement with the experimental. Finally, both the RBS and the HPOTP were modeled and simulated as unstable systems with attached impact dampers. The simulations predict that the impact damper is an excellent stabilizing mechanism for a range of instability driver values.« less