# 参考文献
//基于贝叶斯优化方法，以深度学习训练作业的数据并行形式为基础，预测需要的AWS实例类型和实例数量
@inproceedings{yi2020not,
title={Not All Explorations Are Equal: Harnessing Heterogeneous Profiling Cost for Efficient MLaaS Training},
author={Yi, Jun and Zhang, Chengliang and Wang, Wei and Li, Cheng and Yan, Feng},
booktitle={2020 IEEE International Parallel and Distributed Processing Symposium (IPDPS)},
pages={419--428},
year={2020},
organization={IEEE}
}
// 使用对比学习算法，预测深度学习训练作业在模型并行模式下的资源使用，并给出最优的设备使用结果（同构资源）
@inproceedings{lan2021accelerated,
title={Accelerated Device Placement Optimization with Contrastive Learning},
author={Lan, Hao and Chen, Li and Li, Baochun},
booktitle={50th International Conference on Parallel Processing},
pages={1--10},
year={2021}
}
// Pesto使用算子粒度的建模，估算算子的运行时间，并考虑在模型并行时的算子放置与调度结果
@inproceedings{hafeez2021towards,
title={Towards optimal placement and scheduling of DNN operations with Pesto},
author={Hafeez, Ubaid Ullah and Sun, Xiao and Gandhi, Anshul and Liu, Zhenhua},
booktitle={Proceedings of the 22nd International Middleware Conference},
pages={39--51},
year={2021}
}
// 在云服务之上，如果通过算子预测整个模型的一次运行时间
@inproceedings{hafeez2020empirical,
title={Empirical Analysis and Modeling of Compute Times of CNN Operations on AWS Cloud},
author={Hafeez, Ubaid Ullah and Gandhi, Anshul},
booktitle={2020 IEEE International Symposium on Workload Characterization (IISWC)},
pages={181--192},
year={2020},
organization={IEEE}
}
//机制：nn-Meter将深度学习模型及其算子进行细粒度的划分，能够自动将算子划分为需要的运算单元(由多个算子融合而来)，同时通过数据采样（包含算子的输入数据量和参数维度）便于精准估算算子的运行时间，能够达到接近90%的准确度。NN-meter的局限性在于其只能支持离线融合计算，同时一般算子独占异构资源（边缘环境下有�
