The yin and yang of power and performance for asymmetric hardware and managed software

Pinto

Fundamental Approaches to Software Engineering

2015

Abstract. Empowering application programmers to make energy-aware decisions is a critical dimension of energy optimization for computer systems. In this paper, we study the energy impact of alternative data management choices by programmers, such as data access patterns, data precision choices, and data organization. Second, we attempt to build a bridge between application-level energy management and hardware-level energy management, by elucidating how various application-level data management features respond to Dynamic Voltage and Frequency Scaling (DVFS). Finally, we apply our findings to real-world applications, demonstrating their potential for guiding application-level energy optimization. The empirical study is particularly relevant in the Big Data era, where data-intensive applications are large energy consumers, and their energy efficiency is strongly correlated to how data are maintained and handled in programs.

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Data-Oriented Characterization of Application-Level Energy Optimization

Pinto

Fundamental Approaches to Software Engineering

2015

Lecture Notes in Computer Science

“…Four types of schedulers have been proposed to allocate jobs or parts of jobs to different cores. (1) With known or predicted resource demand, incoming jobs are scheduled to the most appropriate core [9,11,40]. (2) With known performance requirements, latency-sensitive applications such as games or videos are processed by fast cores, whereas latency-tolerant applications such as background services are processed by slow cores [15,25,29].…”

Section: Related Workmentioning

confidence: 99%

“…(1) They must predict demand for each request, scheduling high demand jobs to high performance fast cores and other jobs to low power slow cores [1,9,11,37,40,42]. Unfortunately, the service demand of individual requests in interactive applications is usually unknown and difficult to predict [23].…”

Section: Introductionmentioning

confidence: 99%

A Theoretical Foundation for Scheduling and Designing Heterogeneous Processors for Interactive Applications

Ren

McKinley

2014

To improve performance and meet power constraints, vendors are introducing heterogeneous multicores that combine high performance and low power cores. However, choosing which cores and scheduling applications on them remain open problems. This paper presents a scheduling algorithm that provably minimizes energy on heterogeneous multicores and meets latency constraints for interactive applications, such as search, recommendations, advertisements, and games. Because interactive applications must respond quickly to satisfy users, they impose multiple constraints, including average, tail, and maximum latency. We introduce SEM (Slow-to-fast, Energy optimization for Multiple constraints), which minimizes energy by choosing core speeds and how long to execute jobs on each core. We prove SEM minimizes energy without a priori knowledge of job service demand, satisfies multiple latency constraints simultaneously, and only migrates jobs from slower to faster cores. We address practical concerns of migration overhead and congestion. We prove optimizing energy for average latency requires homogeneous cores, whereas optimizing energy for tail and deadline constraints requires heterogeneous cores. For interactive applications, we create a formal foundation for scheduling and selecting cores in heterogeneous systems.

Proceedings of the 8th ACM International Systems and Storage Conference

“…Therefore, energy is a scarce resource. Previous studies have established that individual components of managed language virtual machines can have a significant impact on energy consumption [Vijaykrishnan et al 2001;Esmaeilzadeh et al 2011;Sartor and Eeckhout 2012;Cao et al 2012;Pinto et al 2014].…”

Section: Introductionmentioning

confidence: 99%

Impact of GC design on power and performance for Android

Hussein

Payer

Hosking

et al. 2015

Small mobile devices have evolved to versatile computing systems. Android devices run a complete software stack, including a full Linux kernel, user land with several software daemons and a virtual machine to run applications. On these mobile systems energy is a scarce resource and needs to be handled carefully. Current systems rely on governors that adjust the frequency of individual cores depending on the system load. We measure energy consumption of different components of this complex software stack, including garbage collection (GC) of the Android virtual machine. Here we propose several extensions to the default GC configuration of Android, including a generational collector, across established dimensions of heap memory size and concurrency.Our evaluation shows that Android's asynchronous GC thread consumes a significant amount of energy. Therefore, varying the GC strategy can reduce total on-chip energy (by 20-30%) whilst slightly impacting application throughput (by 10-40%) and increasing worst-case pause times (by 20-30%). Our work quantifies the direct impact of GC on mobile system, enumerates the main factors and layers of this relationship, and offers a guide for analysis of memory behavior in understanding and tuning system performance.