Investigation into runtime workload classification and management for energy-efficient many-core systems

Abstract

Recent advances in semiconductor technology have facilitated placing many cores on a single chip. This has led to increases in system architecture complexity with diverse application workloads, with single or multiple applications running concurrently. Determining the most energy-efficient system configuration, i.e. the number of parallel threads, their core allocations and operating frequencies, tailored for each kind of workload and application concurrency scenario is extremely challenging because of the multifaceted relationships between these configuration knobs. Modelling and classifying the workloads can greatly simplify the runtime formulation of these relationships, delivering on energy efficiency, which is the key aim of this thesis. This thesis is focused on the development of new models for classifying single- and multi-application workloads in relation to how these workloads depend on the aforementioned system configurations. Underpinning these models, we implement and practically validate low-cost runtime methodologies for energy-efficient many-core processors. This thesis makes four major contributions. Firstly, a comprehensive study is presented that profiles the power consumption and performance characteristics of a multi-threaded many-core system workload, associating power consumption and performance with multiple concurrent applications. These applications are exercised on a heterogeneous platform generating varying system workloads, viz. CPU-intensive or memory-intensive or a combination of both. Fundamental to this study is an investigation of the tradeoffs between inter-application concurrency with performance and power consumption under different system configurations. The second is a novel model-based runtime optimization approach with the aim of achieving maximized power normalized performance considering dynamic variations of workload and application scenarios. Using real experimental measurements on a heterogeneous platform with a number of PARSEC benchmark applications, we study power normalized performance (in terms of IPS/Watt) underpinned with analytical power and performance models, derived through multivariate linear regression (MLR). Using these models we show that CPU intensive applications behave differently in IPS/Watt compared to memory intensive applications in both sequential and concurrent application scenarios. Furthermore, this approach demonstrate that it is possible to continuously adapt system configuration through a per-application runtime optimization algorithm, which can improve the IPS/Watt compared to the existing approach. Runtime overheads vii are at least three cycles for each frequency to determine the control action. To reduce overheads and complexity, a novel model-free runtime optimization approach with the aim of maximizing power-normalized performance considering dynamic workload variations has been proposed. This approach is the third contribution. This approach is based on workload classification. This classification is supported by analysis of data collected from a comprehensive study investigating the tradeoffsbetweeninter-applicationconcurrencywithperformanceand power under different system configurations. Extensive experiments have been carried out on heterogeneous and homogeneous platforms with synthetic and standard benchmark applications to develop the control policies and validate our approach. These experiments show that workload classification into CPU-intensive and memory-intensive types provides the foundation for scalable energy minimization with low complexity. Thefourthcontributioncombinesworkloadclassificationwithmodel based multivariate linear regression. The first approach has been used to reduce the problem complexity, and the second approach has been used for optimization in a reduced decision space using linearregression. This approach further improves IPS/Watt significantly compared to existing approaches. This thesis presents a new runtime governor framework which interfaces runtime management algorithms with system monitors and actuators. This tool is not tied down to the specific control algorithms presented in this thesis and therefore has much wider applications.