Accurate system-level performance modeling and workload characterization for mobile internet devices

Abstract

As mobile applications and devices become ubiquitous, consumer demands for performance, power efficiency, and connectivity are increasing. The software framework existing on mobile internet devices is a complex interaction of real-time tasks, non-real-time applications, and operating system management routines. Traditional simulation approaches are poorly suited to modeling the overall performance characteristics of such systems. Additionally, many traditional benchmark suites used in academia and industry for microprocessor benchmarking and design have been found to be unrepresentative of mobile workloads. This paper presents multiple frameworks utilized for accurately modeling system-level performance of embedded systems used for mobile applications. Furthermore, this paper provides an in-depth workload characterization and memory-level analysis of internet and media-centric applications. All workload characterization is performed running full operating systems and software stacks. For comparison purposes, the system-level analysis is performed at three distinct levels: Full RTL emulation with device peripherals, software-level emulation, and through performance counters on real-world devices. The same mobile workload and operating system is run across all of these platforms.