International Business Machines Corporation

Han Su

Fei Fei Li

Qin Yue Chen

Li Cao

Chao Wei

WeiFeng Zhang

Zhan Peng Huo

Feifei Li

Ren Jie Feng

Zhong Hao Wang

Xin Peng Liu

Yu Sun

Wei Cui

Chang Xin Miao

In a method for improving converter and compiler operator granularity, a processor extracts an operator granularity from an artificial intelligence framework and an original model. A processor receives device characteristics from a target device. A processor outputs a converter granularity level to a converter based on the operator granularity and the device characteristics. A processor outputs a compiler granularity level to a compiler based on the operator granularity and the device characteristics.

Optimizing operator granularity in compiling and converting artificial intelligence (AI) models

Aspects of the invention include converting an artificial intelligence (AI) model generated in a first framework to a uniform exchange formatted model by engaging a master table to retrieve instructions for converting from the AI model to the uniform exchange formatted model in accordance with the first framework. The uniform exchange formatted model in compiled by engaging the master table to retrieve instructions for compiling the uniform exchange formatted model in accordance with the first framework. Data is received as an input to the compiled uniform exchange formatted model and an output is generated by engaging the master table to retrieve instructions for generating the output in accordance with the first framework.

Uniform artificial intelligence model conversion

Embodiments for locating performance hot spots include collecting sample data having instruction addresses, the sample data being for a neural network model and determining instructions in the instruction addresses that are performance hot spots. A listing file is used to map the instructions of the sample data that are performance hot spots to locations in a lower-level intermediate representation. A mapping file is used to map the locations of the lower-level intermediate representation that are performance hot spots to operations in one or more higher-level representations, one or more of the operations corresponding to the performance hot spots, the mapping file being generated from compiling the neural network model.

Locate neural network performance hot spots

Automatic application of software patches to software associated with container images based upon image relationships in a dependency tree. The computing device determines whether software associated with a base container image requires software patches. The computing device accesses dependency trees maintaining image relationships between the base container image and dependent container images. The computing device determines based upon the accessed one or more dependency trees whether the base container image has dependent container images derived from the base container image. The computing device applies software patches to the software associated with the base container image. The computing device rebuilds the base container image with the applied software patches. The computing device then rebuilds the dependent container images dependent upon the rebuilt base container image.

Automatic application of software updates to container images based on dependencies

A method for network communication across application containers in a computer server system includes executing, by a computer system, a host operating system (OS). The host OS is an instance of an OS. The host OS includes multiple application containers operatively coupled to a memory. The method further includes executing, by the host OS, a virtual network interface for each of the application containers. The method further includes implementing, by the host OS, a remote direct memory access (RDMA) transparently for communications amongst the application containers by utilizing shared memory communications.

High performance networking across docker containers

Embodiments of the present disclosure relate to a computer-implemented method for container scheduling in a container orchestration system (COS). According to the method, a new unit comprising one or more containers are detected. Available memory for each of a plurality of candidate nodes deployed in the COS is predicted based on page sharing information of each candidate node. The plurality of candidate nodes filtered to obtain a set of filtered nodes are, wherein the available memory of each of the set of filtered node meets a memory size limitation of the new unit. Priorities of the set of filtered nodes are ranked according to one or more priority functions. The new unit is deployed to one of the filtered nodes based on the priorities.

Container scheduling in a container orchestration system

This disclosure provides methods, systems and computer program products for page sharing among a plurality of containers running on a host. The method comprises in response to a first container accessing a first file not cached by the first container, checking whether a second file equivalent to the first file is shared in a memory of the host by a second container, wherein the checking is based on a record in which related information of at least one shared file is stored. The method further comprises in response to the checking indicating there is no second file, allocating in the memory at least one page for the first file, loading the first file into the at least one page, and storing related information of the first file into the record.

Growing community of inventors

Shanghai, China

Han Su