Workload Uncertainty Characterization and Adaptive Frequency Scaling for Energy Minimization of Embedded Systems

Abstract

Abstract-A primary design optimization objective for multicore embedded systems is to minimize the energy consumption of applications while satisfying their performance requirement. A system-level approach to this problem is to scale the frequency of the processing cores based on the readings obtained from the hardware performance monitors. However, performance monitor readings contain uncertainty, which becomes prominent when applications are executed in a multicore environment. This uncertainty can be attributed to factors such as cache contention and DRAM access time, that are very difficult to predict dynamically. We demonstrate that such uncertainty can be controlled to make better decision on the processor frequency in order to minimize energy consumption. To achieve this, we propose a multinomial logistic regression model, which combines probabilistic interpretation with maximum likelihood (ML) estimation to classify an incoming workload, at run-time, into a finite set of classes. Every workload class corresponds to a frequency pre-determined using an appropriate training set and results in minimum energy consumption. The classifier incorporates (1) uncertainty with arbitrary probability distribution to estimate the actual frame workload; and (2) the frequency switching overhead, neither of which are considered in any of the existing approaches. The classified frequency is applied on the processing cores to execute the workload. The proposed approach is engineered into an embedded multicore system and is validated with a set of standard multimedia applications. Results demonstrate that the proposed approach minimizes energy consumption by an average 20% as compared to the existing techniques.

A Survey of Prediction and Classification Techniques in Multicore Processor Systems

“…The logistic regression based classification is composed of two steps: 1) modeling to estimate the probability distribution of the different classes for a given input, and 2) parameter fitting to estimate the parameters of the logistic regression model. Following the brief description and notation in [34], in the first step, the multinomial logistic regression model works with the assumption that the value of the variable of interest, ∈ [1,2, … , ], is predicted based on the N values of the input feature set, which are identified as = [ 1 , 2 , … , ] ∈ ℝ 1× . The model is represented by the hypothesis ℎ , with parameter ∈ ℝ ( −1)× .…”

“…The study in [34] proposed such a multinomial logistic regression-based classification technique that classifies the workload (i.e., CPU cycles) at runtime into a fixed set of classes. The variable of interest is the workload while specifies the workloads of the previous video frames, where is the workload of the ith previous frame.…”

A railway safety and security concept for low-power mixed-criticality systems

“…In order to reduce static power consumption, the different components can be set in a low power mode or switched-off using power gating technique [11], [12]. With the objective to reduce dynamic power consumption, voltage (also affects static power consumption) [13] and/or frequency scaling [14] and clock gating techniques [15], [16] can be implemented. To manage the implementation of these kind of techniques power mode management techniques can be used [17].…”

Run-time Mapping of Spiking Neural Networks to Neuromorphic Hardware

“…A workload aware thread scheduler is proposed in [20] for multi-processor systems. In [14], the authors propose a multinomial logistic regression model to partition the input workload in run-time. Each partition is then executed at pre-determined frequencies to ensure minimum energy consumption.…”