QoS Driven Coordinated Management of Resources to Save Energy in Multi-Core Systems

Licentiate thesis, 2019

Reducing the energy consumption of computing systems is a necessary endeavor. However, saving energy should not come at the expense of degrading user experience. To this end, in this thesis, we assume that applications running on multi-core processors are associated with a quality-of-service (QoS) target in terms of performance constraints. This way, hardware resources can be throttled to minimize energy expenditure without violating the QoS requirements.

Typical resource management schemes control different resources such as processor cores and on-chip cache memory independently. These approaches are not effective under performance constraints for all applications. Therefore, this thesis presents multi-core resource management schemes that coordinately control several resources in a unified algorithm. This way, the resource manger can find trade-offs between resource allocations to different applications to reduce system-level energy consumption, while still meeting the QoS targets expressed as performance constraints for every application.

Implementing a coordinated resource management scheme that dynamically adapts to varying run time behavior of a multi-programmed workload without any prior knowledge about the applications is a challenging task. Two different schemes are presented in this thesis to address this challenge. Both schemes are invoked at regular intervals during program execution. They employ simple and, yet, sufficiently accurate analytical models and a novel hardware technique to predict the effect of different resource allocations on performance and energy for each application. Using a heuristic method, the multi-dimensional system configuration space is pruned in several levels to find the optimum resource settings, with respect to energy efficiency, in a negligible time.

In the first scheme a resource management algorithm is presented that coordinates the control of voltage-frequency (VF) of each processor core with partitioning of the on-chip cache space. In the second scheme, a re-configurable processor is considered in which sections of the core micro-architectural resources can be dynamically deactivated to save energy. The resource manager can reactivate these sections, at the proper time, to increase instruction and memory level parallelism (ILP/MLP). This introduces new trade-offs between processor core size, VF settings, and the allocation of cache space for each application. By exploiting these trade-offs, the second scheme improves the energy savings compared to the first scheme considerably.

The proposed schemes are evaluated using a novel simulation framework. This framework estimates the effect of different resource management algorithms on full execution of benchmark applications in a multi-programmed workload. According to the experimental results, the proposed schemes can save up to 18% of system energy while respecting the performance constraints of all applications. The average energy savings are 6% and 10% with the first and second schemes, respectively. Further experiments on the first scheme shows that energy savings can potentially improve up to 29% if the users can tolerate a bounded reduction in performance that leads to 40% longer execution time.