Adaptive Task Scheduling and Resource Management Techniques for Improving Energy Efficiency on Multi-core Systems

Doctoral thesis, 2024

The growing impact of energy on operational cost and system robustness becomes a strong motivation for improving energy efficiency in parallel computing systems, in addition to performance. Hardware features such as core asymmetry and Dynamic Voltage and Frequency Scaling (DVFS) aim to provide opportunities for energy-efficient computing. However, it also complicates parallel application development. Task-based parallel programming models have been shown to be a powerful approach for developing parallel applications, allowing developers to express parallelism in the form of tasks. To achieve the goal of energy-efficient execution of a task-based application on multi-core platforms, it is essential to understand application characteristics, underlying platform capabilities and their complex interplay in order to determine appropriate task schedule and resource allocation. Consequently, this thesis introduces four task schedulers - ERASE, STEER, JOSS and SWEEP - tailored for diverse platform capabilities and energy efficiency metrics.

ERASE targets reducing CPU energy consumption in non-user-controllable DVFS environments. The scheduler includes four modules: online performance modeling and power profiling modules provide runtime with execution time and power predictions; core activity tracing offers the instantaneous task parallelism and the task scheduler combines these information to enable CPU energy consumption predictions and dynamically determine the best resource allocation for each task. Moreover, ERASE is designed for quick adaptation to external DVFS changes.

STEER investigates the potential CPU energy savings by leveraging core asymmetry, CPU DVFS, and task characteristics. STEER comprises two predictive models for performance and power predictions, and a task scheduler that utilizes models for energy predictions and then identifies the best resource allocation and frequency settings for tasks. Additionally, STEER employs adaptive scheduling algorithms based on task granularity to handle DVFS overheads and coordinates cluster frequency tuning to mitigate interference from concurrent tasks on cluster-based platforms.

JOSS demonstrates that taking memory energy into account is crucial for reducing total energy consumption, even in the absence of a memory DVFS knob. The scheduler leverages knobs of core asymmetry, CPU DVFS, memory DVFS and task characteristics. JOSS builds a set of models using multivariate polynomial regression, providing predictions for the execution time, average CPU power and memory power of each task, when tuning the aforementioned knobs individually and simultaneously, to facilitate the scheduling decision in task scheduler. Furthermore, JOSS supports exploring energy reduction with and without performance constraints.

SWEEP leverages application attributes, especially inter-task parallelism, together with hardware knobs to predict the impact of task distributions and local task scheduling decisions on the global execution time and energy consumption. SWEEP is designed for exploring various energy performance trade-offs. It first categorizes application execution into high parallelism and low parallelism phases, determined by instantaneous inter-task parallelism. It applies different task scheduling algorithms for high and low parallelism phases respectively, predicting trade-offs associated with different configurations and determining the best task distribution, local task schedules and DVFS settings accordingly.

The four schedulers address the challenges of achieving energy efficiency in diverse computing environments and targeting various energy efficiency metrics for task-based parallel applications. They present a comprehensive approach of integrating predictive models and adaptive scheduling algorithms to fully exploit the capability of multi-core platforms for both energy savings and energy performance trade-offs.