知识卡片
单机最简与分布式复杂度
单机训练是迄今最简单的建模方式,大型云厂商也确实出租配置很强的单台机器;但生产中有意思的 ML 用例往往要处理相当规模的数据,因而可能受益于远不止一台机器。分布式处理带来规模,也带来复杂度——这正是本章后半几乎全部可靠性讨论的根源:多机、多进程、参数状态分散之后,权限、版本、竞态、数据可用性、容量与恢复时间都变成必须显式设计的问题。判断是否要上分布式,应先估算数据量与单机能力,而不是默认训练就该是集群任务。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- Training Systems · 查看原书依据
Training on a single computer is by far the simplest way to build a model, and the large cloud providers do rent powerful configurations of individual machines. Note, though, that many interesting uses of ML in production process a significant amount of data and as a result might benefit from significantly more than one computer. Distributing processing brings scale but also complexity.
参考来源
- 位置:《Reliable Machine Learning》第07章《Training Systems》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch07.xhtml)
- 结论依据:原文明确阐述该主题,并给出定义、实践或反例。
- 原始内容:Training on a single computer is by far the simplest way to build a model, and the large cloud providers do rent powerful configurations of individual machines. Note, though, that many interesting uses of ML in production process a significant amount of data and as a result might benefit from significantly more than one computer. Distributing processing brings scale but also complexity.