Improving the Efficiency of Power Management Techniques by Using Bayesian Classification

Abstract

Abstract1 . This paper presents a supervised learning based dynamic

Learning-Based Run-Time Power and Energy Management of Multi/Many-Core Systems: Current and Future Trends

“…As a result, the classification task is essentially the assignment of the maximum a posteriori (MAP) class given the vector x i and the prior of class assignments to y i [31]. An advantage of naive Bayes is that it only requires a small number of training data to estimate the parameters necessary for classification.…”

“…At run-time, it determines which kernels are likely to best utilize a device from a performance model that predicts a kernel's speedup based on its static code structure. Naive Bayes has also been used in power management to build power-performance model and perform classification [31]. In the context of this work, the goal is to devise a power management policy for issuing DVFS commands on a CMP system that minimize the total energy dissipation based on the load conditions and workload characteristics [37].…”

Learning-Based Power Management for Multicore Processors via Idle Period Manipulation

“…There are also some recent works that consider DPM in multi-core processors, which can be categorized into per-core approach [5,10,11] and chip-wide approach [13,14]. In [10], Canturk et al proposed an approach to set the power mode of each core to meet a power budget.…”

“…In [10], Canturk et al proposed an approach to set the power mode of each core to meet a power budget. Jung et al [5,11] presented a supervised learning-based DPM framework for multi-core processors. Their approach, however, determines power management actions for each core based on their individual workload prediction and hence is not a "true" multi-core power management scheme.…”

“…The effectiveness of the above techniques, however, relies heavily on the accuracy of their workload models, which are not guaranteed during the off-line optimization process. Recently, several learning-based DPM techniques have been presented in the literature (e.g., [4][5][6]12]). By online learning the workload characteristics and adjusting DPM policy on-the-fly, learning-based solutions can adapt to varying system conditions and workloads and hence are potentially able to achieve better power/performance tradeoff than those off-line solutions.…”