To accelerate the training of Deep Learning (DL) models, clusters of machines equipped with hardware accelerators such as GPUs are leveraged to reduce execution time. State-of-the-art resource managers are needed to increase GPU utilization and maximize throughput. While co-locating DL jobs on the same GPU has been shown to be effective, this can incur interference causing slowdown. In this paper we propose Horus: an interference-aware and prediction-based resource manager for DL systems. Horus proactively predicts GPU utilization of heterogeneous DL jobs extrapolated from the DL model's computation graph features, removing the need for online profiling and isolated reserved GPUs. Through micro-benchmarks and job co-location combinations across heterogeneous GPU hardware, we identify GPU utilization as a general proxy metric to determine good placement decisions, in contrast to current approaches which reserve isolated GPUs to perform online profiling and directly measure GPU utilization for each unique submitted job. Our approach promotes high resource utilization and makespan reduction; via real-world experimentation and large-scale trace driven simulation, we demonstrate that Horus outperforms other DL resource managers by up to 61.5% for GPU resource utilization, 23.7-30.7% for makespan reduction and 68.3% in job wait time reduction.
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.
Cloud datacenters capable of provisioning high performance Machine Learning-as-a-Service (MLaaS) at reduced resource cost is achieved via auto-tuning: automated tensor program optimization of Deep Learning models to minimize inference latency within a hardware device. However given the extensive heterogeneity of Deep Learning models, libraries, and hardware devices, performing auto-tuning within Cloud datacenters incurs a significant time, compute resource, and energy cost of which state-of-the-art auto-tuning is not designed to mitigate. In this paper we propose Trimmer, a high performance and cost-efficient Deep Learning auto-tuning framework for Cloud datacenters. Trimmer maximizes DL model performance and tensor program cost-efficiency by preempting tensor program implementations exhibiting poor optimization improvement; and applying an ML-based filtering method to replace expensive low performing tensor programs to provide greater likelihood of selecting low latency tensor programs. Through an empirical study exploring the cost of DL model optimization techniques, our analysis indicates that 26-43% of total energy is expended on measuring tensor program implementations that do not positively contribute towards auto-tuning. Experiment results show that Trimmer achieves high auto-tuning cost-efficiency across different DL models, and reduces auto-tuning energy use by 21.8-40.9% for Cloud clusters whilst achieving DL model latency equivalent to state-of-the-art techniques.
The heterogeneity of Deep Learning models, libraries, and hardware poses an important challenge for improving model inference performance. Auto-tuners address this challenge via automatic tensor program optimization towards a target-device. However, auto-tuners incur a substantial time cost to complete given their design necessitates performing tensor program candidate measurements serially within an isolated target-device to minimize latency measurement inaccuracy. In this paper we propose DOPpler, a parallel auto-tuning measurement infrastructure. DOPpler allows for considerable auto-tuning speedup over conventional approaches whilst maintaining high-quality tensor program optimization. DOPpler accelerates the auto-tuning process by proposing a parallel execution engine to efficiently execute candidate tensor programs in parallel across the CPU-host and GPU target-device, and overcomes measurement inaccuracy by introducing a high-precision on-device measurement technique when measuring tensor program kernel latency. DOPpler is designed to automatically calculate the optimal degree of parallelism to provision fast and accurate auto-tuning for different tensor programs, auto-tuners and target-devices. Experiment results show that DOPpler reduces total auto-tuning time by 50.5% on average whilst achieving optimization gains equivalent to conventional auto-tuning infrastructure.
Federated clusters are composed of multiple independent clusters of machines interconnected by a resource management system, and possess several advantages over centralized cloud datacenter clusters including seamless provisioning of applications across large geographic regions, greater fault tolerance, and increased cluster resource utilization. However, while existing resource management systems for federated clusters are capable of improving application intra-cluster performance, they do not capture inter-cluster performance in their decision making. This is important given federated clusters must execute a wide variety of applications possessing heterogeneous system architectures, which are a impacted by unique inter-cluster performance conditions such as network latency and localized cluster resource contention. In this work we present an empirical study demonstrating how inter-cluster performance conditions negatively impact federated cluster orchestration systems. We conduct a series of micro-benchmarks under various cluster operational scenarios showing the critical importance in capturing inter-cluster performance for resource orchestration in federated clusters. From this benchmark, we determine precise limitations in existing federated orchestration, and highlight key insights to design future orchestration systems. Findings of notable interest entail different application types exhibiting innate performance affinities across various federated cluster operational conditions, and experience substantial performance degradation from even minor increases to latency (8.7x) and resource contention (12.0x) in comparison to centralized cluster architectures.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.