Remote Performance Monitoring (RPM)

As battery-powered, resource-constrained systems continue to grow in capability and complexity, it is increasingly difficult to accurately measure and characterize the full-system power consumption of real devices. However, we must do so if we are to effectively model, predict, and optimize programs and systems to increase battery life. Extant approaches to measurement and characterization of power and energy behavior include simulation, processor-level metrics, and measurement via external monitoring devices (e.g. multi-meters). Hardware performance monitors (HPMs) have gained wide-spread use recently for estimation of CPU processing power [9, 7, 2, 6, 3, 4, 5]. In addition, other types of processor-level metrics have been shown to be effective for predicting CPU performance and power consumption, in particular those related to program phase behavior [8, 1, 4, 5]. These processor-level metrics have been shown in these prior works to correlate well with processor power consumption [2, 3, 4, 5, 6]. Unfortunately, prior work does not evaluate how well processor-level metrics correlate with or estimate the power and energy consumed by the entire system (as opposed to simply the CPU power and energy consumption) or focus on high-end processors such as the Intel Pentium class of processors. We consider Our understanding of full-system energy and of the behavior of energy-efficient processors (e.g. StrongARM and XScale CPUs) is vital if we are to develop techniques for extending the batter life of resource-constrained, mobile devices. To enable the characterization of a full-system as easily and accurately as possible, we developed a toolset called the Remote Performance Monitor (RPM). RPM consists of both hardware and software components. The hardware components include a set of highend tools to monitor the target microcomputer. The software tools include utility programs to configure various system characteristics of the monitored device, and operating system extensions and device drivers to collect performance data (such as HPM counters). A GUI program and a web interface enables remote users (e.g. students and researchers) to submit jobs for performance profiling to our system, i.e, to extract accurate performance profiles without investing in, installing, and managing their own system. RPM collects power, energy, and HPM data for fixedor variable-length intervals. Interval lengths are in terms of dynamic binary instructions and can be set by the user upon job submission. Users of our system can also control which metrics RPM collects and which intervals RPM samples. At present, we use a RPM to characterize a Crossbow Stargate embedded microcomputer. The Stargate implements an Intel XScale processor and a number of I/O devices. The Stargate is very similar in functionality to an HP iPAQ handheld device (without the LCD display), and it is used extensively in sensor network research. In this paper, we employ RPM to investigate how well processorlevel metrics correlate with full-system power and energy consumption by programs. We consider a number of different HPMs as well as a technique that identifies code-based phases in program behavior using simulation. We make many interesting observations using RPM: We find that