Machine Learning on Volatile Instances
Machine Learning on Volatile Instances
复制标题
DOI：
10.1109/infocom41043.2020.9155448
复制
发表时间：
2020-03
期刊：
影响因子：
--
通讯作者：
Xiaoxi Zhang;Jianyu Wang;Gauri Joshi;Carlee Joe-Wong
中科院分区：
文献类型：
--
作者：
Xiaoxi Zhang;Jianyu Wang;Gauri Joshi;Carlee Joe-Wong
由于当今机器学习中使用的神经网络模型和训练数据集规模巨大，因此必须通过在多个工作节点之间分割梯度评估等任务来分配随机梯度下降（SGD）。然而，运行分布式 SGD 可能会非常昂贵，因为它可能需要长时间使用专门的计算资源（例如 GPU）。我们提出了经济高效的策略来利用比标准实例更便宜的易失性云实例，但可能会被更高优先级的工作负载中断。据我们所知，这项工作首次量化了活动工作节点数量的变化（由于抢占）如何影响 SGD 收敛和训练模型的时间。通过了解实例的抢占概率、准确性和训练时间之间的权衡，我们能够得出在易失性实例（例如 Amazon EC2 Spot 实例和其他可抢占云实例）上配置分布式 SGD 作业的实用策略。实验结果表明，我们的策略以显着降低的成本实现了良好的训练性能。
Due to the massive size of the neural network models and training datasets used in machine learning today, it is imperative to distribute stochastic gradient descent (SGD) by splitting up tasks such as gradient evaluation across multiple worker nodes. However, running distributed SGD can be prohibitively expensive because it may require specialized computing resources such as GPUs for extended periods of time. We propose cost-effective strategies to exploit volatile cloud instances that are cheaper than standard instances, but may be interrupted by higher priority workloads. To the best of our knowledge, this work is the first to quantify how variations in the number of active worker nodes (as a result of preemption) affects SGD convergence and the time to train the model. By understanding these trade-offs between preemption probability of the instances, accuracy, and training time, we are able to derive practical strategies for configuring distributed SGD jobs on volatile instances such as Amazon EC2 spot instances and other preemptible cloud instances. Experimental results show that our strategies achieve good training performance at substantially lower cost.