Abstract: Embodiments of this application disclose a data processing method, and relate to the field of artificial intelligence. The method is applied to distributed parallel model training, for example, distributed training of a text translation model, a speech recognition model, a facial recognition model, a three-dimensional reconstruction model, and a virtual reality model. The method can implement hybrid parallelism in a distributed cluster. The method includes: inserting, based on tensor layouts of tensors of at least one operator in a deep neural network model, a redistribution operator between operators that have an input-output dependency relationship, to implement conversion between different tensor layouts; inserting the redistribution operator into a sliced computational graph; and determining an updated sliced computational graph to implement parallel model training of the deep neural network.

Abstract: A container-based processing method includes: when a container state needs to be backed up, determining that content of a first control group file in at least one control group file of the container is modified; acquiring, a file name of the first control group file corresponding to the first flag bit; searching for, according to a correspondence between the file name of the at least one control group file and a storage location of the at least one control group file, a storage location of the first control group file corresponding to the file name of the first control group file, and extracting modified content of the first control group file from the first control group file stored in the storage location of the first control group file; and backing up the file name of the first control group file and the modified content of the first control group file.

Abstract: Embodiments of the present invention provide a method for migrating a user service, and relate to the field of communications, so that a user service is uninterrupted and service data of a user is not lost in a process of migrating an application program. The method includes: saving a user request link message sent by a client; after freezing an application program, saving socket status information corresponding to each current user link, and saving data packets in a socket input buffer and a socket output buffer that correspond to each current user link; and sending the user request link message to a standby server, so that the standby server generates a server socket according to the user request link message; and sending socket status information corresponding to the user link, and the data packets to the standby server, so that the standby server updates the server socket and buffer data.

Abstract: A container-based processing method includes: when a container state needs to be backed up, determining that content of a first control group file in at least one control group file of the container is modified; acquiring, a file name of the first control group file corresponding to the first flag bit; searching for, according to a correspondence between the file name of the at least one control group file and a storage location of the at least one control group file, a storage location of the first control group file corresponding to the file name of the first control group file, and extracting modified content of the first control group file from the first control group file stored in the storage location of the first control group file; and backing up the file name of the first control group file and the modified content of the first control group file.

Abstract: Embodiments of the present invention disclose a method for checkpointing and restarting a container status, including: at a checkpoint stage, acquiring and checkpointing configuration information of a first container, a control group status parameter of the first container, and an application program status parameter of the first container; and at a restart stage, reading the configuration information, the control group status parameter, and the application program status parameter, and configure a second container according to the read configuration information, the read control group status parameter, and the read application program status parameter. The embodiments of the present invention further disclose an apparatus for checkpointing and restarting a container status, so that an execution environment of a process in the container can be kept consistent before and after a checkpoint/restart operation.

Abstract: Embodiments of the present invention provide a method for migrating a user service, and relate to the field of communications, so that a user service is uninterrupted and service data of a user is not lost in a process of migrating an application program. The method includes: saving a user request link message sent by a client; after freezing an application program, saving socket status information corresponding to each current user link, and saving data packets in a socket input buffer and a socket output buffer that correspond to each current user link; and sending the user request link message to a standby server, so that the standby server generates a server socket according to the user request link message; and sending socket status information corresponding to the user link, and the data packets to the standby server, so that the standby server updates the server socket and buffer data.