Optimality of dynamic voltage/frequency scaling in many-core systems with voltage-frequency islands

• Published in: Sustainable computing informatics and systems 2019-12, Vol.24, p.100344, Article 100344
• Main Authors: Hajiamini, Shervin, Shirazi, Behrooz, Dong, Hongbo, Cain, Chris
• Format: Article
• Language: English
• Subjects: Energy budget; Fine-/coarse-grain; ILP; Makespan; MILP; Multicore; Static/Dynamic VFIs
• ISSN: 2210-5379
• DOI: 10.1016/j.suscom.2019.07.007

•Proposing a constrained optimization formulation that solves the task scheduling and V/F level assignment problem for a VFI-based multicore system. This formulation aims to minimize the application task set makespan given an energy budget and tasks’ dependencies.•Solving the above problem for a fine- to coarse-grain VFI-based multicore system. For the fine-grain design wherein each VFI consists of a core, an ILP formulates tasks’ dynamic V/F level assignments given predefined task-core mappings. For the coarse-grain design, two VFI partitioning configurations are proposed, where both use MILP to formulate task scheduling and the static V/F level assignment problems. The first configuration determines VFI partitioning when scheduling tasks onto cores/islands given the tasks’ V/F levels attained from the fine-grain ILP solution. The second configuration optimizes tasks’ V/F levels when scheduling tasks onto a given/predefined VFI partitioned system. For the second configuration, a symmetric VFI-based multicore system is assumed since all the VFIs have an equal number of cores.•To leverage low-cost design of the coarse-grain VFIs and energy efficiency of the fine-grain VFIs (per-core DVFS), this paper presents a two-step optimization method that intelligently forms (asymmetric) VFIs and dynamically performs DVFS per-VFI across the application’s phases. The cores workload behavior is used for partitioning such that the VFIs, with compute-intensive cores, run faster to improve the application’s execution time and the VFIs, which contain cores with less amount of computations, run slower to gain energy saving.•Experimental results over several benchmarks show that the fine-grain and coarse-grain makespans increase with deceasing the energy budget. Furthermore, the fine-grain VFI-based system gains on average 1.35x speedup over the coarse-grain configurations across the experimented energy budgets. Compared to the statically tuned coarse-grain configurations, our DVFS-based VFIs obtain substantial speed up, as much as 1.5x, across the benchmarks. In today’s multicore systems, depending on an application’s computational demand, cores are either operated individually at different Voltage/Frequency (V/F) levels or grouped into multiple Voltage-Frequency Islands (VFIs) to reduce system energy consumption. This paper formulates a task scheduling and VFI partitioning problem whose optimization goal is to minimize the task set (application) execution time (makespan) fo