Energy efficient real-time task scheduling on CPU-GPU hybrid clusters

Mei, Xinxin; Chu, Xiaowen; Liu, Hai; Leung, Yiu-Wing; Li, Zongpeng

doi:10.1109/infocom.2017.8057205

Cited by 61 publications

(39 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…From a technology viewpoint hardware accelerators, such as GPGPUs and FPGAs still need the use of power reduction techniques such as Dynamic Voltage and Frequency Scaling (DVFS) and partial reconfiguration for FPGAs to keep power consumption under control [29]. Many approaches to adaptation and energy/power optimisation concentrate at the hardware level, such as utilising task scheduling coupled with GPU-specific DVFS and dynamic resource sleep (DRS) mechanisms, as a means to minimise the total energy consumption [30]. Our work in this paper compliments such hardware based strategies given the similarity of goals, yet utilises software based approaches to minimise power and conserve energy.…”

Section: Related Workmentioning

confidence: 99%

Energy-Aware Self-Adaptation for Application Execution on Heterogeneous Parallel Architectures

Kavanagh

Djemame

Ejarque

et al. 2020

IEEE Trans. Sustain. Comput.

View full text Add to dashboard Cite

Hardware in High Performance Computing environments in recent years have increasingly become more heterogeneous in order to improve computational performance. An additional aspect of such systems is the management of power and energy consumption. The increase in heterogeneity requires middleware and programming model abstractions to eliminate additional complexities that it brings, while also offering opportunities such as improved power management. In this paper we explore application level self-adaptation including aspects such as automated configuration and deployment of applications to different heterogeneous infrastructure and for their redeployment. This therefore not only mitigates complexities associated with heterogeneous devices but aims to take advantage of the heterogeneity. The overall result of this paper is a self-adaptive framework that manages application Quality of Service (QoS) at runtime, which includes the automatic migration of applications between different accelerated infrastructures. Discussion covers when this migration is appropriate and quantifies the likely benefits.

show abstract

Section: Related Workmentioning

confidence: 99%

Energy-Aware Self-Adaptation for Application Execution on Heterogeneous Parallel Architectures

Kavanagh

Djemame

Ejarque

et al. 2020

IEEE Trans. Sustain. Comput.

View full text Add to dashboard Cite

show abstract

“…Some previous GPU DVFS works indicated that GPUs have more complex energy scaling behaviors, and focused on how to balance the performance and energy efficiency of GPUs [19], [32], [20], [21], [33], [34], [35], [36], [37]. Mei et al [38] and Chau et al [39] further adopted those DVFS-based energy conservation techniques to implement energy-efficient task scheduling for highperformance clusters. Recent papers [40], [41], [42], [43] focused on the performance of scalability of DNN training on different software and hardware environments.…”

Section: B Gpu Dvfsmentioning

confidence: 99%

The Impact of GPU DVFS on the Energy and Performance of Deep Learning

Tang

Wang

et al. 2019

Proceedings of the Tenth ACM International Conference on Future Energy Systems

Self Cite

View full text Add to dashboard Cite

Over the past years, great progress has been made in improving the computing power of general-purpose graphics processing units (GPGPUs), which facilitates the prosperity of deep neural networks (DNNs) in multiple fields like computer vision and natural language processing. A typical DNN training process repeatedly updates tens of millions of parameters, which not only requires huge computing resources but also consumes significant energy. In order to train DNNs in a more energyefficient way, we empirically investigate the impact of GPU Dynamic Voltage and Frequency Scaling (DVFS) on the energy consumption and performance of deep learning. Our experiments cover a wide range of GPU architectures, DVFS settings, and DNN configurations. We observe that, compared to the default core frequency settings of three tested GPUs, the optimal core frequency can help conserve 8.7%∼23.1% energy consumption for different DNN training cases. Regarding the inference, the benefits vary from 19.6%∼26.4%. Our findings suggest that GPU DVFS has great potentials to help develop energy efficient DNN training/inference schemes.

show abstract

“…Any model that considers p‐states and the model calibration process considered here could be combined, enabling greater accuracy. DVFS is also considered in the work of Mei et al in the context of CPU‐GPU hybrid clusters, like in the work of Sundriyal and Sundriyal focusing on scheduling within these heterogeneous environments with power as a key aspect.…”

Section: Related Workmentioning

confidence: 99%

Rapid and accurate energy models through calibration with IPMI and RAPL

Kavanagh

Djemame

2019

Concurrency and Computation

View full text Add to dashboard Cite

Summary Energy consumption in Cloud and High Performance Computing platforms is a significant issue and affects aspects such as the cost of energy and the cooling of the data center. Host level monitoring and prediction provides the groundwork for improving energy efficiency through the placement of workloads. Monitoring must be fast and efficient without unnecessary overhead, to enable scalability. This precludes the use of Watt meters attached per host, requiring alternative approaches such as integrated measurements and models. IPMI and RAPL are subject to error and partial measurement, which may be mitigated. Models allow for prediction and more responsive measures of power consumption, but require calibrating. The causes of calibration error are discussed, along with mitigation strategies, without overly complicating the underlying model. An outcome is a Watt meter emulator that provides hosts level power measurement along with estimated power consumption for a given workload, with an average error of 0.20W.

show abstract

Energy efficient real-time task scheduling on CPU-GPU hybrid clusters

Cited by 61 publications

References 23 publications

Energy-Aware Self-Adaptation for Application Execution on Heterogeneous Parallel Architectures

Energy-Aware Self-Adaptation for Application Execution on Heterogeneous Parallel Architectures

The Impact of GPU DVFS on the Energy and Performance of Deep Learning

Rapid and accurate energy models through calibration with IPMI and RAPL

Contact Info

Product

Resources

About