Statistical power modeling of GPU kernels using performance counters

Nagasaka, Hideki; Maruyama, Naoya; Nukada, Akira; Endo, Tetsuro; Matsuoka, Satoshi

doi:10.1109/greencomp.2010.5598315

Cited by 156 publications

(80 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We use [21] to get the energy consumption of actual program execution and the simulated program execution. Fig.…”

Section: Nonlinear Fitting Of Ga-bp Neural Networkmentioning

confidence: 99%

A Modeling Approach for Energy Saving Based on GA-BP Neural Network

Li¹,

Guo²,

Shen³

et al. 2016

Journal of Electrical Engineering and Technology

View full text Add to dashboard Cite

-To cope with the increasing scale of scientific data and computational complexity of daily data, more and more cores have been integrated into GPU(Graphic Processing Units) and its working frequency is continually upgrading, which makes it being widely used in general computing for assisting CPU to accelerate program. While GPU offers powerful computing capability, the problem of the energy consumption becomes particularly prominently and it has become one of the important issues hindering development of GPU. For the purpose of solving this problem, DVFS (Dynamic Voltage Frequency Scaling) becomes an effective solution. Because the previous works only focus on single component and use linear relationship to do DVFS without considering energy saving of other units in system at software runtime, therefore we propose an energy saving model (CDVFS) of considering the characteristics of both GPU and memory at software runtime based on GA-BP (Genetic Algorithm-Back propagation) neural network to make better use of the relationship between components for energy saving. Firstly, the model assumes that functional relation between the software runtime characteristics of GPU and memory and the appropriate frequency which corresponds to the GPU and memory as nonlinear. Secondly, we extract five characteristics and use GA-BP neural network to fit the nonlinear functional relation. At last, experiments demonstrate the effectiveness of the approach and reasonableness of assumption, and also show that CDVFS can get average energy savings of 17.06% compared with previous works within acceptable performance loss.

show abstract

“…We use [21] to get the energy consumption of actual program execution and the simulated program execution. Fig.…”

Section: Nonlinear Fitting Of Ga-bp Neural Networkmentioning

confidence: 99%

A Modeling Approach for Energy Saving Based on GA-BP Neural Network

Li¹,

Guo²,

Shen³

et al. 2016

Journal of Electrical Engineering and Technology

View full text Add to dashboard Cite

show abstract

“…Exploiting hardware characteristics requires proper models and energy profiles of hardware components. Work in this area abounds: [41] focused on modeling virtual machine's contribution to the consumption of physical nodes; [42,43] proposed statistical power models for GPUs; [44] profiles network equipment. Application scheduling and resource scheduling are emerging now as an integrated approach for large scale distributed systems [45].…”

Section: Existing Research On Energy Efficiencymentioning

confidence: 99%

Power-aware Applications for Scientific Cluster and Distributed

Elmer¹

2014

Proceedings of International Symposium on Grids and Clouds (ISGC) 2014 — PoS(ISGC2014)

View full text Add to dashboard Cite

The aggregate power use of computing hardware is an important cost factor in scientific cluster and distributed computing systems. The Worldwide LHC Computing Grid (WLCG) is a major example of such a distributed computing system, used primarily for high throughput computing (HTC) applications. It has a computing capacity and power consumption rivaling that of the largest supercomputers. The computing capacity required from this system is also expected to grow over the next decade. Optimizing the power utilization and cost of such systems is thus of great interest. A number of trends currently underway will provide new opportunities for power-aware optimizations. We discuss how power-aware software applications and scheduling might be used to reduce power consumption, both as autonomous entities and as part of a (globally) distributed system. As concrete examples of computing centers we provide information on the large HEP-focused Tier-1 at FNAL, and the Tigress High Performance Computing Center at Princeton University, which provides HPC resources in a university context.

show abstract

“…Hotpower [31] is an Nvidia GPU power model based on GPU PMCs, as is the model developed by Nagasaka et al [32]. Wang et al [33] developed a power model based on the PTX assembly code generated by the Nvidia kernel compiler.…”

Section: Related Workmentioning

confidence: 99%

Power Modeling for Heterogeneous Processors

Diop

Jerger

Anderson

2014

Proceedings of Workshop on General Purpose Processing Using GPUs

View full text Add to dashboard Cite

As power becomes an ever more important design consideration, there is a need for accurate power models at all stages of the design process. While power models are available for CPUs and GPUs, only simple models are available for heterogeneous processors. We present a micro-benchmarkbased modeling technique that can be used for chip multiprocessor (CMPs) and accelerated processing units (APUs). We use our approach to model power on an Intel Xeon CPU and an AMD Fusion heterogeneous processor. The resulting error rate for the Xeon's model is below 3% and is only 7% for the Fusion. We also present a method to reduce the number of benchmarks required to create these models. Instead of running micro-benchmarks for every combination of factors (e.g. different operations or memory access patterns), we cluster similar micro-benchmarks to avoid unnecessary simulations. We show that it is possible to eliminate as many as 93% of the compute micro-benchmarks, while still producing power models having less than 10% error rate.

show abstract

Statistical power modeling of GPU kernels using performance counters

Cited by 156 publications

References 15 publications

A Modeling Approach for Energy Saving Based on GA-BP Neural Network

A Modeling Approach for Energy Saving Based on GA-BP Neural Network

Power-aware Applications for Scientific Cluster and Distributed

Power Modeling for Heterogeneous Processors

Contact Info

Product

Resources

About