Techniques to Improve Energy Efficiency on Heterogeneous Multiprocessors under Timing and Quality Constraints

Doctoral thesis, 2022

Traditionally, applications are executed without the notion of a computational deadline and often use all available system resources, which leads to higher energy consumption. User specification of Quality of Service (QoS) constraints, in terms of completion time and solution quality, opens up for allocation of just enough resources to an application to finish just in time and thereby save energy. Modern heterogeneous multiprocessor (HMP) platforms provide a set of configurable resources, including a frequency range of dynamic voltage frequency scaling (DVFS), one among a set processor types, and one or a plurality of processors of each type. They can be configured at run-time to open up new opportunities for resource management.

This thesis presents techniques to reduce energy consumption under QoS constraints by allocating resources at run-time on heterogeneous multiprocessor platforms targeting sequential and parallel iterative and task-parallel applications. The proposed techniques rely on a progress-tracking framework that monitors and predicts how much time is left until the application finishes. Furthermore, the proposed framework enables the prediction of computation demand and performance requirements for future iterations or tasks. The first contribution of this thesis is a resource management technique, called SLOOP, targeting single-threaded applications. SLOOP allocates resources, i.e., processor type and DVFS, for each iteration to meet deadlines while using the prediction of computational demand and execution time.

The second contribution of this thesis is a resource-management scheme, called SaC, for multi-threaded applications executing on HMPs, where resources also include the number of processors besides DVFS and processor type. SaC first chooses the most energy-efficient configuration that meets the deadline. The proposed technique collects execution-time slack over subsequent iterations to select a configuration that can save energy.

The third contribution of this thesis is a resource manager, called Task-RM, for task-parallel applications executing on HMPs under QoS constraints. Task-RM exploits the variance in task execution times and imbalance between sibling tasks to allocate just enough resources in terms of DVFS and processor type. It uses an innovative off-line analysis to avoid redoing scheduling analysis at run-time.

Finally, the fourth contribution is a scheme, called Approx-RM, that can exploit accuracy-energy trade-offs in approximate iterative applications. Approx-RM allocates an appropriate amount of resources while guaranteeing timing and solution quality specifications. Approx-RM first predicts the iteration count required to meet the quality target and then allocates enough resources on an HMP in terms of DVFS, processor type, and processor count to save energy while meeting a performance target.