Enable simultaneous DNN services based on deterministic operator overlap and precise latency prediction
Weihao Cui, Han Zhao, Quan Chen, Ningxin Zheng, Jingwen Leng, Jieru Zhao, Zhuo Song, Tao Ma, Yong Yang, Chao Li, Minyi Guo
摘要
While user-facing services experience diurnal load patterns, co-locating services improve hardware utilization. Prior work on co-locating services on GPUs run queries sequentially, as the latencies of the queries are neither stable nor predictable when running simultaneously. The input sensitiveness and the non-deterministic operator overlap are two primary factors of the latency unpredictability. Hence, We propose Abacus, a runtime system that runs multiple services simultaneously. Abacus enables deterministic operator overlap to enforce latency predictability. Abacus composes of an overlap-aware latency predictor, a headroom-based query controller, and segmental model executors. The predictor predicts the latencies of the deterministic operator overlap. The controller determines the appropriate operator overlap for the QoS guarantee of all the services. The executors run the operators as needed to support the deterministic operator overlap. Our evaluation shows that Abacus reduces 51.3% of the QoS violation and improves the throughput by 29.8% on average compared with state-of-the-art solutions.
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了最后一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
引用它的顶会 Paper15
- Microsecond-scale Preemption for Concurrent GPU-accelerated DNN InferencesMingcong Han, Hanze Zhang, Rong Chen, Haibo ChenOSDI 2022 · 被引用 153 次
- Clover: Toward Sustainable AI with Carbon-Aware Machine Learning Inference ServiceBaolin Li, Siddharth Samsi, Vijay Gadepally, Devesh TiwariSC 2023 · 被引用 63 次
- VELTAIR: towards high-performance multi-tenant deep learning services via adaptive compilation and schedulingZihan Liu, Jingwen Leng, Zhihui Zhang, Quan Chen 等ASPLOS 2022 · 被引用 52 次
- Improving GPU Sharing Performance through Adaptive Bubbleless Spatial-Temporal SharingShulai Zhang, Quan Chen, Weihao Cui, Han Zhao 等EuroSys 2025 · 被引用 19 次
- Efficient Performance-Aware GPU Sharing with Compatibility and Isolation through Kernel Space InterceptionShulai Zhang, Ao Xu, Quan Chen, Han Zhao 等USENIX ATC 2025 · 被引用 16 次
它引用的顶会 Paper11
- Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud ProviderMohammad Shahrad, Rodrigo Fonseca, Iñigo Goiri, Gohar Irfan Chaudhry 等USENIX ATC 2020 · 被引用 946 次
- MLPerf Inference BenchmarkVijay Janapa Reddi, Christine Cheng, David Kanter, Peter Mattson 等ISCA 2020 · 被引用 517 次
- Serving DNNs like Clockwork: Performance Predictability from the Bottom UpArpan Gujarati, Reza Karimi, Safya Alzayat, Wei Hao 等OSDI 2020 · 被引用 392 次
- Sinan: ML-based and QoS-aware resource management for cloud microservicesYanqi Zhang, Weizhe Hua, Zhuangzhuang Zhou, G. Edward Suh 等ASPLOS 2021 · 被引用 226 次
- Rammer: Enabling Holistic Deep Learning Compiler Optimizations with rTasksLingxiao Ma, Zhiqiang Xie, Zhi Yang, Jilong Xue 等OSDI 2020 · 被引用 192 次
相关 Paper
- Tacker: Tensor-CUDA Core Kernel Fusion for Improving the GPU Utilization while Ensuring QoSHan Zhao, Weihao Cui, Quan Chen, Youtao Zhang 等HPCA 2022 · 被引用 42 次
- Orion: Interference-aware, Fine-grained GPU Sharing for ML ApplicationsFoteini Strati, Xianzhe Ma, Ana KlimovicEuroSys 2024 · 被引用 96 次
- Astraea: towards QoS-aware and resource-efficient multi-stage GPU servicesWei Zhang, Quan Chen, Kaihua Fu, Ningxin Zheng 等ASPLOS 2022 · 被引用 28 次
- HSM: A Hybrid Slowdown Model for Multitasking GPUsXia Zhao, Magnus Jahre, Lieven EeckhoutASPLOS 2020 · 被引用 36 次
- Serving Heterogeneous Machine Learning Models on Multi-GPU Servers with Spatio-Temporal SharingSeungbeom Choi, Sunho Lee, Yeonjae Kim, Jongse Park 等USENIX ATC 2022 · 被引用 200 次
