Source-linked AI summary
Machine Learning for Energy-Performance-aware Scheduling
Zheyuan Hu, Yifei Shi
TL;DR
Heterogeneous embedded scheduling must balance energy and latency in a high-dimensional, non-smooth configuration space where heuristic tuning is inefficient. The paper uses Gaussian-Process Bayesian Optimization, Pareto exploration, kernel comparison, and sensitivity analysis to study this problem, reporting non-smoothness, Race-to-Idle, and distinct big- versus little-core roles. The framework is demonstrated for offline parameter tuning, with dynamic environments and dependent tasks left for future work.
Problem
Scheduling heterogeneous processors requires choosing core types and frequencies while balancing high performance against low energy consumption in a complex, non-smooth landscape.
Method
The paper reformulates scheduling parameter tuning as Gaussian-Process Bayesian Optimization with Pareto-frontier exploration, kernel benchmarking, and post-hoc sensitivity analysis.
Results
The study finds that Matérn 5/2 outperforms RBF for sharp performance cliffs, rediscovers Race-to-Idle, and reveals latency roles for big cores alongside energy roles for little cores.
Takeaways & Limitations
The learned model provides interpretable scheduling principles linking heterogeneous hardware resources to energy–latency objectives.
Takeaways & Limitations
The demonstrated framework targets offline parameter tuning; extending it to dynamic environments and dependent tasks remains future work.
Abstract
from arXiv · showhide
In the post-Dennard era, optimizing embedded systems requires navigating complex trade-offs between energy efficiency and latency. Traditional heuristic tuning is often inefficient in such high-dimensional, non-smooth landscapes. In this work, we propose a Bayesian Optimization framework using Gaussian Processes to automate the search for optimal scheduling configurations on heterogeneous multi-core architectures. We explicitly address the multi-objective nature of the problem by approximating the Pareto Frontier between energy and time. Furthermore, by incorporating Sensitivity Analysis (fANOVA) and comparing different covariance kernels (e.g., Matérn vs. RBF), we provide physical interpretability to the black-box model, revealing the dominant hardware parameters driving system performance.
1 Introduction
The paper frames scheduling on heterogeneous processors as a difficult, non-smooth optimization problem and proposes interpretable Bayesian Optimization to explore it. Its analysis highlights kernel suitability and distinct resource roles in energy–latency trade-offs.
- Motivation: Heterogeneous scheduling creates complex, non-linear mappings from configuration choices to energy and throughput, making heuristic tuning and exhaustive search impractical.The configuration space grows combinatorially, while high-fidelity evaluation remains computationally expensive.
- Approach: The framework targets discrete discontinuities in embedded heterogeneous architectures, extending BO beyond the smoother continuous landscapes commonly studied in cloud tuning.The paper investigates whether data-driven modeling can reveal interpretable, physics-compliant scheduling principles.
- Approach: Bayesian Optimization with a Gaussian Process surrogate sequentially balances exploring unsampled configurations against exploiting efficient states.Sensitivity Analysis is integrated into the optimization loop to improve physical interpretability.
- Findings: Matérn 5/2 outperforms the standard RBF kernel by modeling sharp performance cliffs associated with discrete core allocations.This benchmarking result supports a non-smooth characterization of the scheduling landscape.
- Findings: The model autonomously rediscovers Race-to-Idle, with high-frequency big cores often yielding better energy efficiency than leakage-prone low-frequency execution.The reported comparison concerns energy efficiency under the scheduling configurations studied.
- Findings: The optimizer discovers a structural decoupling: latency-critical tasks map to big cores, while little cores support energy conservation.This separates the dominant hardware roles associated with the two objectives.
2 Related works
Related work spans classical schedulers, heterogeneous multi-core systems, and machine-learning-based system optimization. These approaches motivate BO while highlighting challenges in parameter selection, sample efficiency, and interpretability.
- Processor scheduling: Classic FCFS, Round Robin, and priority-based schedulers remain practical, but parameters such as time quantum depend heavily on workload and system settings.Determining suitable scheduler parameters is therefore challenging even with established algorithms.
- Heterogeneous multi-core systems: Heterogeneous systems combine performance-oriented high-frequency cores with energy-efficient low-frequency cores, creating a complex non-smooth parameter space.The Arm big.LITTLE architecture exemplifies this design.
- Machine learning for systems: DRL can address high-dimensional heterogeneous-device mapping, but related work reports poor sample efficiency and limited interpretability.BO and probabilistic programming are presented as more sample-efficient alternatives.
3 Methodology
The methodology models heterogeneous processors, task scheduling, latency, and energy in a simulator, then uses Gaussian Processes and sensitivity analysis to study configuration-performance relationships.
- Simulator design: The simulator represents tasks with arrival time, preferred finish time, priority, instruction count, and energy attributes.Task attributes are initialized over a fixed region, with task arrivals following a Poisson process and exponential waiting times.
- Simulator design: Processor configurations vary across heterogeneous core types and fixed operating frequencies, with DVFS excluded from the model.The methodology investigates how many processor types and what frequencies suit a workload while balancing performance and energy.
- Scheduler design: The simulator implements FCFS, Round Robin, and priority-based scheduling to dispatch tasks across processors.These schedulers differ in ordering, time slicing, fairness, and potential starvation or convoy effects.
- Performance and energy objectives: Latency is modeled from instruction count, fixed IPC, and operating frequency, while priority-aware latency uses larger weights for higher-priority tasks.Turnaround time is the total execution time, and aggregated latency is a weighted average across tasks.
- Performance and energy objectives: Processor energy combines dynamic and leakage power, with frequency-dependent voltage and execution-time relationships defining energy estimates.The model focuses on processor power and separates active, idle, and total energy over execution or fixed time periods.
- Surrogate modeling and interpretation: Gaussian Process regression provides a probabilistic surrogate for the complex scheduling landscape, while anisotropic length scales support post-hoc sensitivity analysis.The covariance kernel encodes smoothness assumptions, and inverse length-scale relationships identify influential input dimensions.
4 Experiments
The experiments evaluate Bayesian Optimization across kernel calibration, preference sensitivity, workload robustness, and Pareto exploration in a mixed-variable scheduling search space.
- Experiment Setup: The framework uses a mixed-variable search space containing continuous, integer, and categorical scheduling parameters.Time quantum is conditional on the selected scheduling strategy.
- Experiment Setup: Each optimization experiment uses 100 trials, beginning with 10 Sobol-sequence warm-up trials to initialize the Gaussian Process surrogate.The Sobol sequence is used for improved space-filling compared with uniform random sampling.
- Surrogate Model Calibration: Kernel calibration compares RBF and Matérn kernels under the standard workload λ = 1.0 to select a baseline for subsequent experiments.The objective is to identify which kernel captures the discrete, non-smooth scheduling landscape more effectively.
- Preference Sensitivity Analysis: Preference sensitivity varies β and γ to represent performance-first, energy-first, and equal-weight objectives.The experiments test whether hardware configuration shifts align with these high-level priorities.
- Workload Robustness Testing: Workload robustness varies the task arrival rate λ from 0.5 to 5.0 to examine how optimal architectural configurations evolve under load.The study specifically examines whether resources scale up to prevent latency cliffs during peak workloads.
- Multi-Objective Pareto Exploration: Multi-objective optimization approximates the Pareto Frontier between energy and time, producing non-dominated solutions without fixed weights.This allows system administrators to make posterior decisions without manually tuning scalarization weights.
5 Results and Analysis
Results show that Matérn kernels, especially Matérn 5/2, better represent the non-smooth scheduling landscape, while sensitivity analysis reveals preference- and workload-dependent hardware strategies. Pareto exploration exposes the energy–time trade-off directly.
- 5.1 Kernel Selection: Matérn 5/2 rapidly identifies the high-performance region within the first 20 trials and maintains low variance afterward.It demonstrates the most stable convergence trajectory among the evaluated kernels.
- 5.1 Kernel Selection: The RBF kernel performs slightly worse because its infinite-smoothness assumption over-smooths discrete scheduling objectives and can miss sharp performance cliffs.The result supports modeling the CPU scheduling landscape as non-smooth.
- 5.1 Kernel Selection: Matérn 3/2 performs comparably to Matérn 5/2 but exhibits slightly higher exploration variance, leading to Matérn 5/2 as the baseline kernel.The selected kernel balances rough-landscape modeling with acquisition-function numerical stability.
- 5.1 Kernel Selection: Sensitivity analysis ranks freq_big_ghz and freq_medium_ghz as dominant under Matérn 5/2, whereas RBF misleadingly ranks freq_little_ghz highest.The authors associate the Matérn ranking with latency sensitivity to high-performance core frequencies.
- 5.2 Metric Evaluation: Time-focused optimization prioritizes freq_big_ghz, quantum_ms, and count_little over count_big, emphasizing peak speed and scheduling granularity.The reported strategy targets reduced preemption overhead rather than simply activating more big cores.
- 5.2 Metric Evaluation: Energy-focused optimization decreases the relative importance of big-core count and instead emphasizes medium-core frequency and little-core count.The resulting strategy is described as efficiency-oriented and associated with offloading toward energy-efficient cores.
- 5.3 Robustness Test: Under λ = 0.5, count_big has Importance ≈0.5 and the optimizer converges toward a minimal little-core configuration.The reported configuration deactivates big cores to minimize static power while maintaining acceptable latency.
- 5.3 Robustness Test: At λ = 2.5, the optimizer activates 4 × Little, 4 × Medium, and 4 × Big cores, then focuses on freq_medium_ghz.This operating point balances latency and energy by tuning the frequency of power-efficient clusters.
6 Conclusion
The framework reformulates scheduling parameter tuning as Bayesian Optimization and uses sensitivity and multi-objective analyses to reveal physical behavior in post-Dennard architectures. Its offline scope leaves extensions for dynamic runtime adaptation and dependent tasks.
- Bayesian Optimization reformulates heterogeneous multi-core scheduling parameter tuning while sensitivity and multi-objective analyses expose physical system behavior.
- The Matérn 5/2 kernel outperforms RBF because the scheduling landscape contains sharp performance cliffs rather than smooth gradients.
- The optimizer validates Race-to-Idle in energy-constrained scenarios and decouples latency control to big cores from energy management by little cores.
- Optimal strategy changes at saturation: resource scaling is effective for λ ≤2.5, whereas λ = 5.0 triggers penalty mitigation that minimizes energy.
- The demonstrated framework remains an offline configuration method, motivating future work on online runtime adaptation rather than fixed schedules.
- The current assumption of independent tasks is targeted for relaxation through DAG-based dependency models addressing communication overhead and pipeline stalling.
A Details and additional results
Additional results include Bayesian Optimization histories organized by metric type and by lambda rate. These figures support comparisons across the corresponding experimental dimensions.
- Figure 9 presents Bayesian Optimization history results for different types of metrics.
- Figure 10 presents Bayesian Optimization history results for different lambda rates.
B Proposal
The proposal uses machine learning to learn how CPU scheduling affects the efficiency–energy trade-off. It targets influential scheduling behavior, tunable algorithms, and interpretable decision support.
- The proposal uses machine learning to learn how CPU scheduling affects the efficiency–energy trade-off.
- It investigates which scheduling behaviors, including pre-emption frequency, load balance, and task affinity, most influence efficiency and energy use.
- It examines how conventional or hybrid algorithms can be tuned to optimize throughput, latency, and power consumption.
- It studies how machine learning can provide scheduling insights by learning parameters, predicting workloads, or guiding real-time decisions.
C Author contributions
The authors jointly discussed and researched relevant prior work, assisted with debugging, and participated in reviewing and improving the overall codebase.
- Both authors jointly discussed and researched prior work, assisted each other with debugging, and reviewed and improved the overall codebase.
Zheyuan (zh369)
This work implemented scheduling simulation frameworks and task models while contributing to the system’s Gaussian Process, kernel, Bayesian Optimization, and Sensitivity Analysis components.
- Implementation: SimPy frameworks implemented FCFS, time-sliced, Round Robin, and priority-based scheduling algorithms.
- Implementation: Task models were designed and implemented for the project’s scheduling simulations.
- Implementation: Scheduler logging tracked task attributes, assignments, core energy consumption, latency, fairness, and throughput.
- Optimization: The optimization component included Gaussian Processes, kernel choices, Bayesian Optimization, and Sensitivity Analysis.
- Writing and editing: Contributions included authoring the Abstract, Related Work, and Simulation sections and editing figures, tables, and results.
Yifei (ys690)
This work implemented optimization from simulator outputs and contributed to scheduling simulators and task models. It also acknowledges feedback received during the project’s proposal and viva phases.
- Optimization: The optimization component used simulator outputs for latency, throughput, fairness, and energy.
- Methodology: The overall methodology covered Gaussian Processes, kernel choices, Bayesian Optimization, and Sensitivity Analysis.
- Scheduling: Scheduling simulator contributions included FIFO, time-quantum-based, Round Robin, and priority-based scheduling.
- Task modeling: Task models were designed and developed as part of the scheduling simulator work.
- Acknowledgement: Professor Carl Henrik Ek provided feedback during the project’s proposal and viva phase.