上海科技大学知识管理系统

本发明提供一种基于非易失内存的深度神经网络checkpoint优化系统以及方法，通过客户端模块以及服务端模块在深度神经网络训练开始前将对应的网络结构注册在非易失内存中，并创建数据索引和基于远程直接内存访问(RDMA)的数据通信协议；并且在神经网络训练过程中，本发明提供了零拷贝、异步、端到端的神经网络数据持久化，使得用户可以在不影响训练速度的前提下做细粒度的checkpointing以保证容错性和数据持久性。

Provided in the present invention are a deep neural network checkpoint optimization system and method based on a nonvolatile memory. A corresponding network structure is registered in a nonvolatile memory by means of a client module and a serving end module before deep neural network training is started, and a data index and a data communication protocol based on remote direct memory access (RDMA) are created; and during a neural network training process, zero-copy asynchronous end-to-end neural network data persistence is provided in the present invention, such that a user can do fine-grained checkpointing without affecting a training speed, thus ensuring fault tolerance and data persistence.