Trading-off Accuracy vs Energy in Multicore Processors via Evolutionary Algorithms Combining Loop Perforation and Static Analysis-Based Scheduling

Abstract

Abstract. This work addresses the problem of energy efficient scheduling and allocation of tasks in multicore environments, where the tasks can permit certain loss in accuracy of either final or intermediate results, while still providing proper functionality. Loss in accuracy is usually obtained with techniques that decrease computational load, which can result in significant energy savings. To this end, in this work we use the loop perforation technique that transforms loops to execute a subset of their iterations, and integrate it in our existing optimisation tool for energy efficient scheduling in multicore environments based on evolutionary algorithms and static analysis for estimating energy consumption of different schedules. The approach is designed for multicore XMOS chips, but it can be adapted to any multicore environment with slight changes. The experiments conducted on a case study in different scenarios show that our new scheduler enhanced with loop perforation improves the previous one, achieving significant energy savings (3f % on average) for acceptable levels of accuracy loss.

An evolutionary scheduling approach for trading-off accuracy vs. verifiable energy in multicore processors

“…To approximate upper and lower bounds on the energy consumption of a schedule, we use the approach that we proposed in [3]. It combines (dynamic) energy modelling techniques, to infer energy bounds on the program's basic blocks by using an EA, with static analysis techniques, that use these bounds on basic blocks to infer bounds on the whole program as a function of its input data sizes.…”

“…Our previous work [3] used average energy models, which give an average energy consumption of each task in a schedule and hence an average energy consumption of the whole schedule. Herein we extend our energy-efficient scheduling algorithm by using upper-and lower-bound energy models that provide safe upper and lower bounds on the energy consumption of each task and hence on the whole schedule.…”

ENTRA: Whole-systems energy transparency

“…EAs have also been used to improve energy-aware allocation and scheduling for DVFS-enabled multicore environments. For example, the algorithms described in [46,47] are able to deal with task migration and preemption; and the ones in [48] allow decreasing program accuracy (by performing loop perforation) in order to save energy.…”