A power‐efficient scheduling methodology leveraging thread‐based parallel programming to optimize resource utilization across CPU–GPU architectures using the taskflow framework and a dynamic programming algorithm is introduced to optimize computing task assignment and computational resource allocation.
Abstract
CPU–GPU heterogeneous computing platforms are increasingly ubiquitous in mobile embedded systems, including smartphones, tablets, and unmanned aerial vehicles, due to their superior computational throughput. However, given the strict energy limitations inherent to battery‐powered devices, the development of power‐efficient, real‐time computing architectures remains a critical design challenge. This paper proposes a power‐efficient scheduling methodology leveraging thread‐based parallel programming to optimize resource utilization across CPU–GPU architectures using the taskflow framework. By explicitly modeling thread‐level runtime behavior across heterogeneous computation units, the scheduling objective is formulated as an integer nonlinear programming problem aimed at minimizing total system power consumption. A state‐transition framework is derived, and a dynamic programming algorithm is introduced to optimize computing task assignment and computational resource allocation. Extensive empirical evaluations demonstrate that the proposed dynamic programming strategy significantly accelerates convergence and yields superior power‐efficient schedules compared to baseline methodologies, including particle swarm optimization. Furthermore, validation on a physical platform confirms the practical efficacy and accuracy of the proposed scheduling architecture.
DEFT is presented, an energy-aware scheduling framework that jointly optimizes task-to-device assignment and per-GPU DVFS configuration for task-based multi-GPU applications and employs a cost-model–driven strategy that integrates slack awareness, throughput awareness, and explicit modeling of task execution cost, inter-GPU data movement, and DVFS transition overheads, enabling coordinated placement and frequency decisions under dynamic runtime conditions.
Jing Chen, Miquel Pericàs· International Conference on...· 0 citations
Agentic AI systems compose heterogeneous tool workloads on shared GPU/CPU infrastructure, yet existing frameworks assign all GPU-capable tools to the GPU by default. We profile 19 AI tools across GPU and CPU and find that 11 are GPU-preferred, 4 are ambiguous, 1 is CPU-preferred due to PCIe transfer dominance, and 3 are device-neutral, establishing that blanket GPU-first scheduling is suboptimal. We formulate device scheduling as assigning each tool to one of three options: immediate GPU execution, queued GPU execution, or CPU offload, under a VRAM budget, and identify two runtime factors that cause end-to-end latency to diverge from static profiles: GPU utilization contention and VRAM capacity contention. We present an agentic scheduler that pairs an LLM agent with an algorithmic runtime monitor, where the monitor expands what the LLM can observe via running averages, symmetric reprobing, swap reprobing, and exploration hints, without ever prescribing which mapping to adopt. Across 13 scenarios spanning serial execution, parallel contention, and memory-constrained execution, the agentic scheduler reaches the brute-force optimal mapping in all 13 scenarios, matching the best classical baseline on mapping accuracy while avoiding bandit-style exploration over complete mappings, and outperforming HEFT, StarPU, and the all-GPU policy while requiring zero offline training.
Accurate and scalable switching power analysis remains a critical bottleneck in modern physical design, often forcing a trade-off between computational speed and modeling fidelity. We present DiffPower, a GPU-accelerated framework for differentiable power analysis and optimization. DiffPower translates design netlists into a PDK-agnostic bytecode representation, enabling analytical gradient computation via reverse-mode automatic differentiation, achieving up to a $1{,}002\times$ speedup over single-threaded CPU propagation on the largest evaluated design, with the GPU advantage growing with design scale. A hybrid propagation methodology fusing analytical modeling with parallel simulation achieves a median toggle-rate correlation of $r{=}0.96$ across ten industrial and benchmark designs. The resulting \emph{power gradients}, computed up to $904\times$ faster than CPU finite-difference methods with near-perfect rank agreement, enable two downstream applications: (1) gradient-weighted cell sizing, which achieves up to $2.98\times$ improvement over local-power heuristics on industrial designs, with even stronger advantages at the 117K-cell scale where competing methods plateau; and (2) power virus generation via gradient ascent, which yields up to $2.13\times$ higher transition-weighted power, replacing a search process that traditionally requires hours.
Isaac Jacobson, Zhengjie Zhao, R.C. Mehrotra et al.· 0 citations
Energy efficiency has become a primary design constraint in modern high-performance computing systems, where energy costs increasingly dominate total cost of ownership. At the same time, applications are subject to strict performance requirements, making energy optimization inherently a multi-objective problem involving conflicting objectives of minimizing energy consumption while preserving application performance [1, 2]. Hardware interfaces such as Intel RAPL enable runtime power capping, allowing dynamic control over processor power budgets, but selecting optimal operating points remains challenging due to application-dependent and non-linear performance behavior [5, 7]. This work presents Decay-Driven Multi-Objective Reinforcement Learning (DDMORL), an offline preference-driven reinforcement learning framework for adaptive power control in HPC systems. The proposed approach extends preference-driven multi-objective reinforcement learning [3] to a fully offline setting [9], eliminating the need for unsafe online exploration and enabling deployment in production environments. Unlike prior approaches that rely on fixed scalarization or multiple trained policies, DDMORL learns a single preference-conditioned controller that spans the entire energy performance trade-off space. A key contribution of this work is a decay-driven preference formulation that replaces abstract preference weights with a user-defined performance degradation parameter. Users specify a maximum tolerable decay, which is then analytically mapped to a preference vector aligned with the corresponding operating point on the Pareto front. This mapping leverages a domain-specific monotonic relationship between application progress and power cap, removing the need for interpolation models typically required in preference-driven MORL.
Akhilesh Raj, Swann Perarnau, A. Gokhale· IEEE International Symposium...· 0 citations
Cloud computing has transformed the delivery of modern applications and services by providing scalable, flexible, and cost-effective access to computing resources. One of the most critical challenges in cloud environments is the efficient distribution of dynamic workloads across heterogeneous resources, commonly addressed through load balancing and task scheduling techniques. Efficient scheduling plays a vital role in maximizing resource utilization, minimizing response time, and maintaining acceptable Quality of Service (QoS), particularly under dynamic and large-scale workloads. Despite the progress achieved by traditional heuristics such as Min-Min and metaheuristic approaches like the Improved Sparrow Search Algorithm (ISSA), challenges related to scalability, adaptability, and computational overhead remain. Metaheuristic-based approaches often involve iterative optimization processes that may limit their efficiency in real-time scheduling scenarios. In this paper, we propose a lightweight Stochastic Predictive Energy-Aware Scheduling (SPES) algorithm that integrates predictive execution estimation, multi-resource awareness, and stochastic decision-making. Unlike deterministic scheduling strategies, SPES employs a Top K candidate selection mechanism combined with probabilistic weighting and epsilon-greedy exploration to enhance adaptability and avoid suboptimal resource allocation. The proposed method considers CPU, memory, and I/O demands to achieve balanced utilization across heterogeneous hosts while implicitly addressing energy efficiency through utilization-based modeling. The proposed algorithm is implemented and evaluated using the CloudSim 5.0 simulation framework under heterogeneous multi-region cloud environments with varying workload sizes. Experimental results demonstrate that SPES consistently outperforms ISSA and achieves makespan reductions of up to 23.8% while improving scalability, resource utilization, and scheduling efficiency under dynamic cloud workloads. These results indicate that SPES provides an effective lightweight scheduling solution for large-scale and energy-aware cloud computing environments and supports green computing objectives through improved resource efficiency.
M. Yacoub, Ahmed E. Abdel Raouf, Walaa K. Gad et al.· Electronics· 0 citations
DySM is a closed-loop response time control algorithm for spatially shared GPUs in soft real-time systems designed using feedback control theory for guaranteed system stability and control accuracy and outperforms state-of-the-art solutions by providing runtime real-time guarantees.
Srinivasan Subramaniyan, Xiaorui Wang· Euromicro Conference on Real...· 0 citations