System-level integrated power management for handheld systems

Abstract

We propose a system-level integrated power management scheme for battery-operated handheld systems such as cell phones and PDAs. Rather than dealing separately with each system component, we consider the interactions between CPU, WNIC (wireless network interface card), LCD, and applications, to reduce energy consumption at the system-level. Depending on the type of applications, the proposed scheme takes the interaction between CPU voltage and frequency and either LCD clock frequency or WNIC power modes, selectively, or both of them. The proposed method selects voltage for CPU in the context of LCD clock speed to reduce the system energy consumption. The application type and the power mode of WNIC are also considered to control the CPU voltage and frequency. Experimental results show that our scheme reduces the system energy consumption by as much as 30% compared to the systems of simply combining DVS (dynamic voltage scaling) and DPM (dynamic power management) or those of using no energy saving policy.

Introduction

Modern handheld systems such as PDAs (personal data assistants), PMPs (portable media players), or cell phones tend to run complex applications and typically require Internet connectivity. Power management is important for handheld devices because of their limited battery capacity. In general, a handheld system is composed of CPU, memory, WNIC (wireless network interface card), LCD, and system bus. In this paper we focus on the CPU, WNIC, and LCD because they are the main power consumers. CPU may take 6–65% of the total power, LCD 12–65%, and WNIC 4–35% in handheld systems [1], [2]. In particular, the energy consumption of LCDs has been growing as the size increases as well as the increasing popularity of multimedia applications, games, and DMB (digital multimedia broadcasting) applications.

Many power management techniques for handheld systems have been proposed and they can be categorized as either component-level or system-level method. Component-level power management aims to reduce the energy consumption of individual components, whereas system-level power management reduces that of the entire system. Most of the previous researches on power management have considered CPU, WNIC, and LCD separately, on the assumption that the power consumption of each component is orthogonal. These methods simply combine individual techniques for each component into a power management policy for the overall system [5], [11], [16], [18], [26], [27], [28].

System-level power management methods take into account the main system components to reduce the system energy consumption [3], [4], [5]. The previous approaches can be classified as network-CPU methods, battery-CPU methods, memory-CPU methods, LCD–CPU methods, and others. Network-CPU methods control the CPU frequency according to the status of the network packet queues or network information [6], [7], [8], [9], [10]. Battery-CPU methods manage the CPU frequency depending on the battery residual and draining velocity [11], [12], [13], [14], [15]. Memory-CPU methods adjust the CPU frequency to extend the CPU idle period when system executes memory-bound jobs [16], [17]. The relationship between energy consumption and clock frequencies for a microprocessor and a memory was considered in [17]. The methods in [11], [18] control the clock frequency used for processor-display communication, which reduces the LCD energy consumption by decreasing the refresh rate.

To develop an effective way of system-level power management mechanism, we consider the interactions among main components, CPU, WNIC, and LCD. Several efforts classify applications for managing a WNIC. When the CPU frequency is controlled, the proposed method decides the frequency scaling step in the context of the LCD clock frequency. We heuristically found the control factors to save energy at the system-level by conducting preliminary experiments to observe the interactions of the system components. Based on the preliminary results we propose a system-level integrated power management scheme. The mechanism uses the interaction between the CPU voltage and frequency and the LCD clock frequency, and includes our prior work which considered the interaction between the CPU voltage and frequency and the WNIC power modes according to the type of applications [19], [20]. In this paper, we use network applications for workload to consider the CPU, LCD, and WNIC together. If it is not easy to control CPU, LCD, and WNIC in harmony in terms of integrated management, each pair for integrated management, CPU–LCD or CPU–WNIC, has still effect of reducing energy consumption at system-level. For example, if the application not using network is executed, the CPU–WNIC method cannot be applied. In this case, the proposed CPU–LCD method can be applied and reduce energy consumption of the system.

This paper analyzes the interactions between the CPU and LCD as well as WNIC, and then proposes an integrated power management scheme, called S-IPM (system-level integrated power management). S-IPM classifies incoming network traffics by workload type at the kernel level, which has been updated by modifications of the previous work [21]. It can control the CPU voltage and frequency and the WNIC power mode to suit a specific application in the QoS (Quality of Service) context. The mechanism scales the CPU voltage and frequency with the knowledge that the LCD controller’s clocks will be affected by the CPU frequency. S-IPM checks the running applications periodically with about 100 ms interval in this paper. Handheld systems characteristically place weight on accessing a single application because of a performance or battery problem. One drawback of the proposed method is a periodic checking of the application type, which may slightly reduce the system performance. However, this process is a common mechanism to schedule processes in general purpose operating systems such as Linux.

We have implemented the S-IPM mechanism on both the Mainstone II and the Zaurus SL-C860 PDA. The Mainstone II is equipped with PXA270 processor and a TFT LCD display, and runs the Linux operating system. The Zaurus SL-C860 has PXA255 processor and also runs Linux. For workloads, we use three types of applications: ftp, streaming, and web application. The efficiency of S-IPM is shown by the Energy × Delay product, which takes into account QoS. Experimental results show that S-IPM reduces the system energy consumption by 12–23%, on the average, compared to both no energy-saving policy and the conventional method that simply combines DVS and DPM.

The rest of this paper is organized as follows. In Section 2 we discuss related work. Section 3 describes the motivation and principles behind our approach. In Section 4 we discuss experimental results. Section 5 concludes the paper.