Energy-Efficient Real-Time Scheduling of DAG Tasks

Bhuiyan, Ashikahmed; Guo, Zhishan; Saifullah, Abusayeed; Guan, Nan; Xiong, Haoyi

doi:10.1145/3241049

Cited by 49 publications

(16 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The reason for this behaviour is due to convolution kernels being intrinsically more compute intensive in contrast to GEMM kernels, as well as different memory transfer and compute patterns when contending for resources [6]. 4 The Horus Framework…”

Section: Discussionmentioning

confidence: 99%

“…DL jobs are assigned resources and allocated onto machines through use of a resource manager to increase system resource efficiency to satisfy a specified Service Level Agreement (SLA). Existing DL system resource managers 3 have focused on a specific sub-set of objectives including minimizing makespan, JCT, as well as maximizing system resource utilization and energy-efficiency [4].…”

Section: Deep Learning Resource Managersmentioning

confidence: 99%

“…E(GU til j ). Such features are well known and frequently modified by Machine Learning researchers and developers [27,33,32], hence it is relatively straight forward for practitioners to manually extract, or automatically collect such metrics via graph analysis tools such as TensorFlow profiler 4 and TorchScript 5 . Estimation of expected GPU Memory: The only exception for straight forward metric collection the total job memory size (MiB) due to initialization and optimization of individual DL libraries.…”

Section: Resource Estimatormentioning

confidence: 99%

See 2 more Smart Citations

Horus: An Interference-Aware Resource Manager for Deep Learning Systems

Yeung

Borowiec

Yang

et al. 2020

Algorithms and Architectures for Parallel Processing

View full text Add to dashboard Cite

Deep Learning (DL) models are deployed as jobs within machines containing GPUs. These DL systems-ranging from a singular GPU device to machine clusters-require state-of-the-art resource management to increase resource utilization and job throughput. While it has been identified that co-location-multiple jobs co-located within the same GPU-is an effective means to achieve this, such co-location incurs performance interference that directly debilitates DL training and inference performance. Existing approaches to mitigate interference require resource intensive and time consuming kernel profiling ill-suited for runtime scheduling decisions. Current DL system resource management are not designed to deal with these problems. This paper proposes Horus, an interference-aware resource manager for DL systems. Instead of leveraging expensive kernel-profiling, our approach estimates job resource utilization and co-location patterns to determine effective DL job placement to minimize likelihood of interference, as well as improve system resource utilization and makespan. Our analysis shows that interference cause up to 3.2x DL job slowdown. We integrated our approach within the Kubernetes resource manager, and conduct experiments in a DL cluster by training 2,500 DL jobs using 13 different models types. Results demonstrate that Horus is able to outperform other DL resource managers by up to 61.5% for resource utilization and 33.6% for makespan.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Deep Learning Resource Managersmentioning

confidence: 99%

Section: Resource Estimatormentioning

confidence: 99%

See 1 more Smart Citation

Horus: An Interference-Aware Resource Manager for Deep Learning Systems

Yeung

Borowiec

Yang

et al. 2020

Algorithms and Architectures for Parallel Processing

View full text Add to dashboard Cite

show abstract

“…To date, a significant amount of works studied the energy minimization scheduling technique considering both the parallel and sequential real-time tasks, in a non-MC platform, few to mention [11], [12], [15], [17], [26]- [28], [41]. On the other hand, extensive research has been done on real-time scheduling of the MC task model considering both the sequential and parallel workload model (e.g., [4]- [6], [14], [20], [29], [34]), without considering the energy-awareness.…”

Section: Related Workmentioning

confidence: 99%

Optimizing Energy in Non-Preemptive Mixed-Criticality Scheduling by Exploiting Probabilistic Information

Bhuiyan

Reghenzani

Fornaciari

et al. 2020

IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst.

Self Cite

View full text Add to dashboard Cite

The strict requirements on the timing correctness biased the modeling and analysis of real-time systems towards the worst-case performances. Such focus on the worst-case, however, does not provide enough information to effectively steer the resource/energy optimization. In this paper, we integrate a probabilistic-based energy prediction strategy with the precise scheduling of mixed-criticality tasks, where the timing correctness must be met for all tasks at all scenarios. The Dynamic Voltage and Frequency Scaling (DVFS) is applied to this precise scheduling policy to enable energy minimization. We propose a probabilistic technique to derive an energy-efficient speed (for the processor) that minimizes the average energy consumption, while guaranteeing the (worst-case) timing correctness for all tasks, including lo-criticality ones, under any execution condition. We present a response time analysis for such systems under the nonpreemptive fixed-priority scheduling policy. Finally, we conduct an extensive simulation campaign based on randomly generated task sets to verify the effectiveness of our algorithm (w.r.t. energy savings) and it reports up to 46% energy-saving.

show abstract

“…Although a significant amount of works studied the energyaware real-time task scheduling in the non-MC platform (few to mention [10,11,23,24,33,34]), little progress has been made on energy-aware scheduling in the MC platform. Huang et al [27] proposed an energy-aware scheduling by increasing the processor speed after a mode-switch.…”

Section: Related Workmentioning

confidence: 99%