Machine Learning-Based Approaches for Energy-Efficiency Prediction and Scheduling in Composite Cores Architectures

Sayadi, Hossein; Patel, Nisarg; Sasan, Avesta; Homayoun, Houman

doi:10.1109/iccd.2017.28

Cited by 40 publications

(15 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…In contrast to the previous approaches, recent work have proposed models that can predict what core configuration will best meet an application's requirements [32,40]. [1] proposed an offline approach to optimize energy and performance using data mining to determine an optimal scheduling of applications to cores.…”

Section: Related Workmentioning

confidence: 99%

“…[1] proposed an offline approach to optimize energy and performance using data mining to determine an optimal scheduling of applications to cores. [32] used several machine learning algorithms to determine an energy-efficient scheduling of applications to cores, where the work focused on different core architectures in contrast to our work which focuses on different cache configurations. [42] used a metric known as Performance Impact Estimation (PIE) to predict the performance of different workloads when executed on different cores for making decisions on whether to migrate a specific workload to a different core.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Dynamic Scheduling for Heterogeneous Multicores

et al. 2021

View full text Add to dashboard Cite

Heterogeneous multicore systems can help adherence to design goals by providing a diverse set of hardware components to meet application requirements. Each core may also have tunable hardware that can reconfigured for different applications. However, scheduling becomes difficult in the presence of tunable hardware due to the additional constraint that an application must be scheduled to a core that offers the best configuration, based on the application's requirements, to maximize the benefit of the tunable hardware. The scheduling decision can be made by exploring and evaluating each hardware configuration in the design space exhaustively or through the use of a heuristic to evaluate a subset of the design space. However, exhaustive and heuristic approaches do not scale well with the number of available hardware configurations. To improve upon exhaustive and heuristic approaches, in this paper, we present a dynamic scheduling methodology that uses machine learning to schedule applications to the application's respective best core for reduced energy consumption for a system with configurable caches. We use an artificial neural network (ANN) to train our predictive model, using hardware counters, to predict the best core and a tuning heuristic to determine best configuration on non-best cores. If the best core is busy, our scheduler considers alternative idle cores, or the application is stalled depending on which decision is energy advantageous. Our experiments show that our proposed system can achieve a total energy savings of approximately 33% and 2.5% compared to a fixed-cache, base system when using a configurable instruction and data cache, respectively, while remaining competitive and exhibiting better scaling properties compared to systems adopting a heuristic approach.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Dynamic Scheduling for Heterogeneous Multicores

et al. 2021

View full text Add to dashboard Cite

show abstract

“…In [27], the authors apply machine learning to find out energy-efficient con- Table 1, and thus need to implement some non-trivial module to collect such data at runtime. On the other hand, these studies alleviate the performance prediction problem of mappings by either focusing on task/thread executions on some specific resources such as in [21,25,26,24] without considering the communication aspects, or focusing on the prediction of threads and/or cores numbers or core configurations such as in [23,22,27] without investigating the explicit thread/task-core binding solutions. No microarchitecture-dependent information is required in our approach contrarily to approaches such as [29] or [30].…”

Section: Related Workmentioning

confidence: 99%

Empirical model-based performance prediction for application mapping on multicore architectures

Gamatié

Zhang

et al. 2019

Journal of Systems Architecture

View full text Add to dashboard Cite

Application mapping in multicore embedded systems plays a central role in their energy-efficiency. The present paper deals with this issue by focusing on the prediction of performance and energy consumption, induced by task and data allocation on computing resources. It proposes a solution by answering three fundamental questions as follows: i) how to encode mappings for training performance prediction models? ii) how to define an adequate criterion for assessing the quality of mapping performance predictors? and iii) which technique among regression and classification enables the best predictions? Here, the prediction models are obtained by applying carefully selected supervised machine learning techniques on raw data, generated off-line from system executions. These techniques are Support Vector Machines, Adaptive Boosting (AdaBoost) and Artificial Neural Networks (ANNs). Our study is validated on an automotive application case study. The experimental results show that with a limited set of training information, AdaBoost and ANNs can provide very good outcomes (up to 84.8% and 89.05% correct prediction score in some cases, respectively), making them attractive enough for the addressed problem.

show abstract

“…manager to dynamically tailor the hardware configuration to the changing program execution characteristics. It avoids the pitfalls of a static optimization strategy where the hardware configuration remains unchanged for the dynamically evolving program phases [23], [24], [25], [26].…”

Section: Introductionmentioning

confidence: 99%

“…Unlike prior machine-learning-based approaches [23], [24], [25], [26], [32], where the learned model remains static after deployment, we use RL to continually refine and update the decision policy throughout runtime execution. As the RL system learns and adjusts its decisions over time, it gains a better understanding of what works for the running program and becomes more efficient in recommending hardware power configurations for the target programs and underlying hardware.…”

Section: Introductionmentioning

confidence: 99%

Online Power Management for Multi-Cores: A Reinforcement Learning Based Approach

Wang

Zhang

Hao

et al. 2022

IEEE Trans. Parallel Distrib. Syst.

View full text Add to dashboard Cite

Power and energy is the first-class design constraint for multi-core processors and is a limiting factor for future-generation supercomputers. While modern processor design provides a wide range of mechanisms for power and energy optimization, it remains unclear how software can make the best use of them. This paper presents a novel approach for runtime power optimization on modern multi-core systems. Our policy combines power capping and uncore frequency scaling to match the hardware power profile to the dynamically changing program behavior at runtime. We achieve this by employing reinforcement learning (RL) to automatically explore the energy-performance optimization space from training programs, learning the subtle relationships between the hardware power profile, the program characteristics, power consumption and program running times. Our RL framework then uses the learned knowledge to adapt the chip's power budget and uncore frequency to match the changing program phases for any new, previously unseen program. We evaluate our approach on two computing clusters by applying our techniques to 11 parallel programs that were not seen by our RL framework at the training stage. Experimental results show that our approach can reduce the system-level energy consumption by 12%, on average, with less than 3% of slowdown on the application performance. By lowering the uncore frequency to leave more energy budget to allow the processor cores to run at a higher frequency, our approach can reduce the energy consumption by up to 17% while improving the application performance by 5% for specific workloads.

show abstract

Machine Learning-Based Approaches for Energy-Efficiency Prediction and Scheduling in Composite Cores Architectures

Cited by 40 publications

References 19 publications

Dynamic Scheduling for Heterogeneous Multicores

Dynamic Scheduling for Heterogeneous Multicores

Empirical model-based performance prediction for application mapping on multicore architectures

Online Power Management for Multi-Cores: A Reinforcement Learning Based Approach

Contact Info

Product

Resources

About