Unlock instant, AI-driven research and patent intelligence for your innovation.

TensorFlow distributed matrix calculation implementation method and system

A technology of matrix calculation and implementation method, which is applied in the field of distributed matrix calculation implementation method and system, which can solve problems such as difficult to use and weak support of general-purpose graphics cards, and achieve the effect of improving computing performance

Active Publication Date: 2021-03-16
SHANGHAI JIAO TONG UNIV
View PDF5 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

Compared with some commonly used servers at present, ARM servers have relatively weak support for general-purpose graphics cards. In the field of machine learning, NVIDIA GPU applications and developer ecology are relatively more complete, and ARM-based servers are currently difficult to use these tools

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • TensorFlow distributed matrix calculation implementation method and system
  • TensorFlow distributed matrix calculation implementation method and system
  • TensorFlow distributed matrix calculation implementation method and system

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0022] This embodiment relates to a distributed matrix computing system for TensorFlow on an ARM server, that is, a cluster, which is composed of server nodes participating in computing, one of which is a master node, and runs a TensorFlow program and a computing agent program on the master node , only run the distribution handler on the remaining server nodes that are compute nodes.

[0023] The master node refers to: a node in the cluster, the node runs the TensorFlow program and the calculation agent program.

[0024] The calculation agent program communicates with TensorFlow to obtain matrix data and returns calculation results, and at the same time communicates with the distribution processing program to distribute matrix data, receive calculation results, and perform subsequent processing on the calculation results.

[0025] The distributed processing program runs in all non-main node servers, receives matrix data from the calculation agent program for calculation, and r...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a TensorFlow distributed matrix calculation implementation method and system, and the method comprises the steps: intercepting an underlying matrix multiplication calculation operation during the operation of TensorFlow, transmitting a calculation parameter to a calculation agent program, carrying out the distributed processing, achieving the transmission of data between different processes through employing a shared memory, and achieving the distributed matrix calculation. The method and system do not depend on specific computer hardware such as a high-performance GPU(Graphics Processing Unit), does not need to modify and adapt an operation program, and can enable TensorFlow codes to perform multi-node parallel operation by directly performing distributed processing on bottom-layer calculation by utilizing a plurality of ARM servers.

Description

technical field [0001] The present invention relates to a technology in the field of distributed information processing, in particular to a distributed matrix calculation implementation method and system for TensorFlow on an ARM server. Background technique [0002] Under the direction of large-scale and distributed operation, the current multi-machine parallel operation of machine learning models can be divided into two types: data parallelism and model parallelism. Among them, data parallelism is a widely used method. Data parallel computing acceleration refers to performing the same computing process on different data on multiple computing nodes to obtain the final result. This parallel method is widely used in multi-GPU parallelism and multi-node parallelism. It can distribute data sets on different nodes to support large data sets, and can also utilize the computing power of various computing devices and computing nodes. [0003] Parallel operation at the model level ...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06F9/54G06F17/16G06N3/04G06N3/08G06N20/00
CPCG06F17/16G06F9/544G06N20/00G06N3/08G06N3/045Y02D10/00
Inventor 郑文立李亭君沈耀陈全过敏意
Owner SHANGHAI JIAO TONG UNIV