User- and Process-Driven Dynamic Voltage and Frequency Scaling Bin Lin Arindam Mallik Peter Dinda Gokhan Memik Robert Dick {b-lin,arindam,pdinda,g-memik,dickrp}@northwestern.edu Department of EECS, Northwestern University Abstract existing DVFS techniques in high-performance processors select an operating point (CPU frequency and voltage) based on the We describe and evaluate two new, independently-applicable utilization of the processor. While this approach can integrate in- power reduction techniques for power management on proces- formation available to the OS kernel, such control is pessimistic. sors that support dynamic voltage and frequency scaling (DVFS): Existing DVFS techniques are pessimistic about the user. user-driven frequency scaling (UDFS) and process-driven volt- They assume that CPU utilization or the OS events prompting age scaling (PDVS). In PDVS, a CPU-customized profile is de- it are sufficient proxies. A high CPU utilization simply leads to rived offline that encodes the minimum voltage needed to achieve a high frequency and high voltage, regardless of the user’s satis- stability at each combination of CPU frequency and tempera- faction or expectation of performance. ture. On a typical processor, PDVS reduces the voltage below Existing DVFS techniques are often pessimistic about the the worst-case minimum operating voltages given in datasheets. CPU. They assume worst-case manufacturing process variation UDFS, on the other hand, dynamically adapts CPU frequency to and operating temperature by basing their policies on loose the individual user and the workload through direct user feed- worst-case bounds given by the processor manufacturer. A volt- back. Our UDFS algorithms dramatically reduce typical operat- age level for each frequency is set such that even the slowest ing frequencies and voltages while maintaining performance at shipped processor of a given generation will be stable at the high- a satisfactory level for each user. We evaluate our techniques est specified temperature. independently and together through user studies conducted on a In response to these observations, on which we elaborate in Pentium M laptop running Windows applications. We measure Sections 2.1 and 3.1, we have developed two new power man- the overall system power and temperature reduction achieved by agement techniques that can be readily employed independently our methods. Combining PDVS and the best UDFS scheme re- or together. In particular, we introduce the following techniques. duces measured system power by 49.9% (27.8% PDVS, 22.1% User-Driven Frequency Scaling (UDFS) uses direct user feed- UDFS), averaged across all our users and applications, com- back to drive an online control algorithm that determines the pro- pared to Windows XP DVFS. The average temperature of the ◦ cessor frequency (Section 2.2). Processor frequency has strong CPU is decreased by 13.2 C. User trace-driven simulation to effects on power consumption and temperature, both directly and evaluate the CPU only indicates average CPU dynamic power also indirectly through the need for higher voltages at higher fre- savings of 57.3% (32.4% PDVS, 24.9% UDFS), with a maxi- quencies. The choice of frequency is directly visible to the user as mum reduction of 83.4%. In a multitasking environment, the it determines observed performance. There is considerable vari- same UDFS+PDVS technique reduces the CPU dynamic power ation among users with respect to the satisfactory performance by 75.7% on average. level for a given workload mix. UDFS exploits this variation to customize frequency control policies dynamically to the individ- 1 Introduction ual user. Unlike previous work (Section 5), we employ direct feedback from the user during ordinary use of the machine. Dynamic Voltage and Frequency Scaling (DVFS) is one of Process-Driven Voltage Scaling (PDVS) creates a custom the most commonly used power reduction techniques in high- mapping from frequency and temperature to the minimum volt- performance processors and is an important OS power manage- age needed for CPU stability (Section 3.2), taking advantage of ment tool. DVFS is generally implemented in the kernel and it process variation. This mapping is then used online to choose varies the frequency and voltage of a microprocessor in real-time the operating voltage by taking into account the current oper- according to processing needs. Although there are different ver- ating temperature and frequency. Researchers have shown that sions of DVFS, at its core DVFS adapts power consumption and process variation causes IC speed to vary up to 30% [2]. Hence, performance to the current workload of the CPU. Specifically, using a single supply voltage setting does not exploit the variation This work is in part supported by DOE Awards DE-FG02-05ER25691 and DE- in timing present among processors. Although some processors AC05-00OR22725 (via ORNL), NSF Awards CNS-0720691, CNS-0721978, customize voltage–frequency mappings based on process varia- CNS-0715612, CNS-0551639, CNS-0347941, CCF-0541337, CCF-0444405, tion, none adjust voltage as a function of temperature. PDVS CCF-0747201, IIS-0536994, IIS-0613568, ANI-0093221, ANI-0301108, and EIA-0224449, by SRC award 2007-HJ-1593, by Wissner-Slivka Chair funds, and does both. We take advantage of the variation among ICs via a by gifts from Symantec, Dell, and VMware. customization process that determines the slack of the individual processor, as well as its dependence on operating temperature. 2 User-Driven Frequency Scaling (UDFS) This offline measurement is used online to dynamically set volt- Current DVFS techniques are pessimistic about the user, age based on frequency and temperature. which leads them to often use higher frequencies than necessary We evaluate our techniques independently and together for satisfactory performance. In this section, we elaborate on through user studies conducted on a Pentium M laptop run- this pessimism and then explain our response to it: user-driven ning Windows applications. Our studies, described in detail in frequency scaling (UDFS). Evaluations of UDFS algorithms are Section 4, include both single task and multitasking scenarios. given in Section 4. We measure the overall system power and temperature reduc- tion achieved by our methods. Combining PDVS and the best 2.1 Pessimism about the user UDFS scheme reduces measured system power by 49.9% (27.8% Current software that drives DVFS is pessimistic about the in- PDVS, 22.1% UDFS), averaged across all our users and applica- dividual user’s reaction to the slowdown that may occur when tions, compared to the Windows XP DVFS scheme. The average CPU frequency is reduced. Typically, the frequency is tightly ◦ temperature of the CPU is decreased by 13.2 C on average. Us- tied to CPU usage. A burst of computation due to, for example, ing user trace-driven simulation to evaluate the CPU in isolation, a mouse or keyboard event brings utilization quickly up to 100% we find average CPU dynamic power savings of 57.3% (32.4% and drives frequency, voltage, temperature, and power consump- PDVS, 24.9% UDFS), with a maximum reduction of 83.4%. In tion up along with it. CPU-intensive applications also cause an a multitasking environment, the same UDFS+PDVS technique almost instant increase in operating frequency and voltage. reduces the CPU dynamic power by 75.7% on average. In both cases, the CPU utilization (or OS events that drive it) 1.1 Experimental setup is functioning as a proxy for user comfort. Is it a good proxy? To find out, we conducted a small (n = 8) randomized user Our experiments were done using an IBM Thinkpad T43p study, comparing four processor frequency strategies including with a 2.13 GHz Pentium M-770 CPU and 1 GB memory run- dynamic, static low frequency (1.06 GHz), static medium fre- ning Microsoft Windows XP Professional SP2. Although eight quency (1.33 GHz), as well as static high frequency (1.86 GHz). different frequency levels can be set on the Pentium M-770 pro- The dynamic strategy is the default DVFS policy used in Win- cessor, only six can be used due to limitations in the SpeedStep dows XP Professional. Note that the processor maximum fre- technology. quency is 2.13 GHz. We allowed the users to acclimate to the full In all of our studies, we make use of three application tasks, speed performance of the machine and its applications for 4 min- some of which are CPU intensive and some of which frequently utes and then carry out the tasks described in Section 1.1, with block while waiting for user input: the following durations: • Creating a presentation using Microsoft PowerPoint 2003 • PowerPoint (4 minutes in total, 1 minute per strategy) while listening to background music using Windows Media • Shockwave (80 seconds in total, 20 seconds per strategy) Player 10. The user duplicates a presentation consisting of • FIFA (4 minutes in total, 1 minute per strategy) complex diagrams involving drawing and labeling, starting Users verbally ranked their comfort levels after each task / strat- from a hardcopy of a sample presentation. egy pair on a scale of 1 (discomfort) to 10 (very comfortable). • Watching a 3D Shockwave animation using the Microsoft Note that for each application and user, strategies were tested in Internet Explorer web browser. The user watches the random order. animation and is encouraged to press the number keys to Figure 1 illustrates the results of the study in the form of over- change the camera’s viewpoint. The animation was stored lapped histograms of the participants’ reported comfort level for locally. Shockwave options were configured so that each of four strategies. Consider Figure 1(a), which shows re- rendering was done entirely in software on the CPU. sults for the PowerPoint task. The horizontal axis displays the • Playing the FIFA 2005 Soccer game. FIFA 2005 is a range of comfort levels allowed in the study and the vertical axis popular and widely-used First Person Shooter game.