Investigation into scalable energy and performance models for many-core systems

Abstract

It is likely that many-core processor systems will continue to penetrate emerging embedded and high-performance applications. Scalable energy and performance models are two critical aspects that provide insights into the conflicting trade-offs between them with growing hardware and software complexity. Traditional performance models, such as Amdahl’s Law, Gustafson’s and Sun-Ni’s, have helped the research community and industry to better understand the system performance bounds with given processing resources, which is otherwise known as speedup. However, these models and their existing extensions have limited applicability for energy and/or performance-driven system optimization in practical systems. For instance, these are typically based on software characteristics, assuming ideal and homogeneous hardware platforms or limited forms of processor heterogeneity. In addition, the measurement of speedup and parallelization factors of an application running on a specific hardware platform require instrumenting the original software codes. Indeed, practical speedup and parallelizability models of application workloads running on modern heterogeneous hardware are critical for energy and performance models, as they can be used to inform design and control decisions with an aim to improve system throughput and energy efficiency. This thesis addresses the limitations by firstly developing novel and scalable speedup and energy consumption models based on a more general representation of heterogeneity, referred to as the normal form heterogeneity. A method is developed whereby standard performance counters found in modern many-core platforms can be used to derive speedup, and therefore the parallelizability of the software, without instrumenting applications. This extends the usability of the new models to scenarios where the parallelizability of software is unknown, leading to potentially Run-Time Management (RTM) speedup and/or energy efficiency optimization. The models and optimization methods presented in this thesis are validated through extensive experimentation, by running a number of different applications in wide-ranging concurrency scenarios on a number of different homogeneous and heterogeneous Multi/Many Core Processor (M/MCP) systems. These include homogeneous and heterogeneous architectures and viii range from existing off-the-shelf platforms to potential future system extensions. The practical use of these models and methods is demonstrated through real examples such as studying the effectiveness of the system load balancer. The models and methodologies proposed in this thesis provide guidance to a new opportunities for improving the energy efficiency of M/MCP systems