Mathematical Problems in Engineering

Volume 2015, Article ID 815230, 12 pages

http://dx.doi.org/10.1155/2015/815230

## A Formal Approach for RT-DVS Algorithms Evaluation Based on Statistical Model Checking

College of Computer Science, Sichuan University, Chengdu 610065, China

Received 1 June 2015; Accepted 12 July 2015

Academic Editor: Xiaoyu Song

Copyright © 2015 Shengxin Dai et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

Energy saving is a crucial concern in embedded real time systems. Many RT-DVS algorithms have been proposed to save energy while preserving deadline guarantees. This paper presents a novel approach to evaluate RT-DVS algorithms using statistical model checking. A scalable framework is proposed for RT-DVS algorithms evaluation, in which the relevant components are modeled as stochastic timed automata, and the evaluation metrics including utilization bound, energy efficiency, battery awareness, and temperature awareness are expressed as statistical queries. Evaluation of these metrics is performed by verifying the corresponding queries using UPPAAL-SMC and analyzing the statistical information provided by the tool. We demonstrate the applicability of our framework via a case study of five classical RT-DVS algorithms.

#### 1. Introduction

Energy saving has become a crucial concern for embedded real time systems, especially for the battery-powered systems. High energy consumption not only shortens the lifespan of computing system but also leads to high processor temperature that degrades system performance, robustness, and reliability. Therefore, system level power management has gained significant attention in recent years.

Dynamic voltage scaling (DVS) is a way to save energy in processor, which adjusts the supply voltage and operating frequency dynamically. The energy consumption scales quadratically to the supply voltage (), and the operating frequency almost linearly depends on the supply voltage. DVS permits trading system performance for energy saving through frequency and voltage scaling [1]. However, changing the operating frequency will affect the execution time of the tasks and may violate the timing constraints. Real time DVS (RT-DVS) [1] is proposed to achieve energy saving while ensuring temporal correctness.

A number of RT-DVS algorithms [1–4] have been developed in the last two decades. Many of these algorithms focus on periodic tasks, assuming that the tasks timing attributes including period, worst-case execution time (*WCET*, at the maximum available processor frequency), and deadline are known in advance. These algorithms have been experimentally studied using simulation based approaches both on software simulators [5, 6] and on hardware platforms [7, 8]. However, simulation based evaluation is time consuming and often not completely reliable. Hence, we propose evaluating DVS algorithms with statistical model checking approach which is based on mathematical models and logics.

In this paper, we illustrate the applicability of statistical model checking [9], an automatic formal verification technique, to the area of RT-DVS algorithms evaluation. The relevant components involved in RT-DVS algorithms evaluation are modeled as stochastic timed automata [10] and implemented in statistical model checker UPPAAL-SMC [10]. Subsequently, the evaluation metrics including utilization bound, energy efficiency, battery awareness, and temperature awareness are expressed as statistical queries and written in metric interval temporal logic (MITL) [10]. Finally, the system model is verified with respect to statistical queries using UPPAAL-SMC. The model checker estimates the probability of the system to satisfy the queries. What is more, the tool provides statistical information, for example, probability densities and means, which is used for further evaluation. We propose a scalable framework for RT-DVS algorithms evaluation. In the sequel, a case study is conducted in order to demonstrate the applicability of our framework.

To the best of our knowledge, this is the first research effort that proposes evaluating RT-DVS algorithms using statistical model checking. Our conclusions are partly observed by the simulation based approaches. However, we obtain these conclusions in a more automatic and reliable manner. What is more, we consider the impact of temperature on the energy efficiency of RT-DVS algorithms. In addition, it should be emphasized that our approach is not limited to the algorithms described in this paper, but all the dynamic voltage scaling algorithms can be modeled and then evaluated with our framework. Furthermore, although multicore systems are increasingly used, single core systems are still of interest since the approaches can be applied to multicore systems with some modification.

The rest of the paper is structured as follows. Section 2 introduces the fundamental theories of RT-DVS algorithms. Section 3 introduces statistical model checking based approach for RT-DVS algorithms evaluation. Section 4 describes the system model in detail. Section 5 describes the statistical queries for evaluation. Section 6 presents a case study to evaluate five classical RT-DVS algorithms. Section 7 gives an overview of the related work. Finally, Section 8 concludes the paper.

#### 2. Power Consumption and RT-DVS Algorithms

This section introduces some fundamental theories of RT-DVS algorithms, which are the foundation of the models presented later.

##### 2.1. Processor Power Distribution

The power consumption of the processor can be divided into three components: dynamic, static, and short circuit power [4]. Short circuit power is comparatively insignificant to dynamic and static power; thus it is negligible [4], so we do not discuss it in this paper.

A DVS enabled processor has a discrete set of operating frequencies , and each clock frequency is associated with a supply voltage level in set . Meanwhile processor has two modes of operation: idle mode when it does not have any job in the system and thus dissipates only static power and active mode when the processor executes some jobs in the system and dissipates both dynamic power and static power [11]. For many years, dynamic power dissipation was the dominant factor in total power consumption which can be approximated with the following formula:where is the load capacitance, is the average number of circuit switches per clock, is the supply voltage, and is the operating frequency. The operating frequency almost linearly depends on the supply voltage [4]:where is a constant and is the threshold voltage. By substituting (2) in (1), the dynamic power at supply voltage will be formulated as [12]where is a processor specific constants and is th voltage level.

Static power or leakage power increases its dominance of total power consumption in recent years, which can be formulated as [11]where represents the number of gates, is the leakage current, and has a nonlinear exponential relationship with both temperature and supply voltage. Fortunately, it has been shown in [13] that the static power can be approximatively calculated withwhere and are some processor specific voltage sensitive constants and is the processor temperature. As reported in [13], using (5) can accurately estimate the static power, with relative error less than 1.3%.

In general, the total power of the processor at active model is formulated as

Lastly, we discuss the processor temperature in detail. We adopt the well-known lumped RC model [12] to describe the transient behavior of the processor temperature aswhere and are processor specific thermal coefficients and and are the instantaneous power and temperature, respectively. Without loss of generality, we have scaled the temperature such that the ambient temperature is considered zero. Otherwise, an offset equals the ambient temperature which is substituted in (7).

##### 2.2. RT-DVS Algorithms

RT-DVS algorithms exploit two types of slack time [3]: static slack, when the total utilization is less than 1, there exists time interval that can be used for voltage scaling and dynamic slack, when the actual execution time of the task is smaller than its* WCET*, a time interval is generated to further reduce the supply voltage. RT-DVS algorithms differ in the way they estimate and utilize the static and dynamic slacks; accordingly, they can be further divided into three types as follows:(1)We have static voltage scaling [1], for example, static Earliest Deadline First (*staEDF*) and static Rate Monotonic (*staRM*), which only exploits the static slacks by setting the operating frequency to the lowest possible one that guarantees the taskset feasibility. Note that the frequency is set at the starting and never changed during execution.(2)We have cycle conserving scaling [1], for example, cycle conserving EDF (*ccEDF*) and cycle conserving RM (*ccRM*), which exploits the dynamic slacks by updating the utilization whenever a task is released or terminated; thus when a task is terminated early, the remaining time can be used to scale the frequency down.(3)We have look-ahead scaling [1], for example, look-ahead EDF (*laEDF*), which exploits both static and dynamic slacks by pushing tasks as close as possible to their deadlines and sets the operating frequency to the slowest possible one that ensures that all future deadlines are met.

#### 3. RT-DVS Algorithms Evaluation Based on Statistical Model Checking

##### 3.1. Statistical Model Checking and UPPAAL-SMC

Statistical model checking [9] is one of the quantitative verification [14] approaches, which efficiently addresses verification problem by combining the Monte Carlo method with temporal logic model checking [15]. It automatically samples the traces (or simulations) of the system model, checks whether the samples satisfy or violate the quantitative property, and finally applies a statistical estimation technique to compute an approximate value for the probability that the property is satisfied [15]. The estimation is correct up to some confidence level that can be parameterized by user.

UPPAAL-SMC [10] is a stochastic and statistical model checking extension of UPPAAL toolbox for verification of stochastic hybrid systems. The modeling formalism of UPPAAL is an extension of classical timed automata with additional features such as integer variables, data structures, user-defined functions, and synchronization. And the modeling formalism of UPPAAL-SMC is based on a stochastic interpretation and extension of the timed automata used in UPPAAL; it has been extended to handle stochastic and nonlinear dynamic behaviors, discrete probabilities, a stochastic interpretation for timed delays, and even dynamic process creation [10]. In UPPAAL-SMC, a system is represented as a network of interacting stochastic timed automata (STAs), which communicate via broadcast channels and shared variables [10]. Three types of queries are supported by UPPAAL-SMC, that is, probability estimation, hypothesis testing, and probability comparison. The probability confidence interval can be estimated by the query , the hypothesis testing is achieved by the query (where is a specified threshold), and probability comparison is achieved by the query , where defines how to bound the runs. There are three ways to bound the runs, they are by time (specifying where is positive integer), by cost (specifying where is a specific clock), and by number of steps (specifying ) [10]. In addition to the three classical queries, UPPAAL-SMC supports the evaluation of expected values of min or max of an expression that evaluates to a clock or an integer value; the queries are or , respectively, where is as previously explained, accounts for the number of runs explicitly, and is the expression to be evaluated.

##### 3.2. Statistical Model Checking Based Approach for RT-DVS Algorithms

Our approach is illustrated in Figure 1. Each RT-DVS algorithm is modeled as timed automata. Tasksets generator model consists of tasksets attributes generator model and task template. It is used to generate the tasksets to which the algorithms can be applied. Platform model consists of processor model and battery model, in which the variation of energy consumption, processor temperature, and the electronic charges in battery is modeled according to the analytical models proposed in existing papers, such as [3, 12, 13, 16]. UPPAAL-SMC automatically composites the models in advance; then the entire system model for evaluation can be obtained.