Maximizing Utilization and Minimizing Migration in Thermal-Aware Energy-Efficient Real-Time Multiprocessor Scheduling

Abstract

This work proposes CAlECs, a clustered scheduling system for MPSoCs subject to thermal and energy constraints. It calculates off-line a cyclic executive honoring temporal and thermal constraints, for a hard real-time (HRT) task set at minimum frequency to reduce consumed energy, minimizing context switches and migrations. It also provides an on-line controller able to manage system and task parametric variations and soft real-time (SRT) tasks, always meeting the HRT task set constraints and the system thermal bound. CAlECS maximizes CPU utilization to help avoid overprovisioning contributing to a low SWaP factor. Its modular design allows the utilization of different modeling and scheduling approaches, and makes the off-line and on-line components independent from each other to better suit the requirements of a specific system. We experimentally show that the cyclic executive provided by CAlECS for HRT task sets outperforms RUN, a reference off-line algorithm in terms of optimal number of context switches.

Accounting for Preemption and Migration Costs in the Calculation of Hard Real-Time Cyclic Executives for MPSoCs

“…AdWCET can virtually use any appropriate scheduler to compute a CE at each iteration. This section summarizes CAlECS [2], the scheduler we leverage in this paper because of its properties (Sec. I).…”

“…2 shows CALEC's modules. The task set conditioner receives the task set T and processor set P, and outputs a subset F ⊆ F of frequencies such that the task set is feasible, thermal safe and energy efficient [2]. Then, the task clustering module tries to partition the task set into subsets (Q k ) that could fit in clusters of sizes s j ≤ m. If the cluster has size equal to one, then it is scheduled using EDF [21], otherwise the cluster is scheduled using AlECS [22], which is an strategy that encompasses a deadline partitioned scheme, a linear programming problem and a zero-laxity policy.…”

“…Algorithms for RT multiprocessor scheduling can be broadly classified into partitioned, global, and hybrid schemes (semipartitioned, clustered) [1], [2]. Partitioned schedulers [3], [4] statically allocate tasks to processors.…”

“…Recent research shows that we can get best of two worlds by calculating a predictable, low-overhead CE while obtaining the optimal utilization provided by global or clustered schemes [2], [13]. However, there remain two non-negligible problems to solve.…”

“…Novelty As far as we know, this paper introduces the first method to improve the calculation of CEs considering the migration preemption costs, providing a more realistic path to obtain predictable HRT CEs with optimal utilization and minimum overhead on a multicore architecture.The proposed methodology can be applied to any multiprocessor RT preemptive scheduler ( [13]) but in this work we consider implicit-deadline HRT task set and leverage CAlECS [2]. This hybrid RT scheduler optimizes processor utilization at the lowest possible energy to comply thermal and temporal constraints, obtaining a CE with very few preemptions and migrations, but obviates the aggregated overhead.…”