Neural Network Layer Data Partitioning for Overlapped Node Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The serial processing of data in trillion-level large models across communication nodes is time-consuming, as the entire data is sent, processed, and then the result is sent back, leading to inefficiencies in data processing time.
Innovation Solution
The data is divided into sub-data pieces, which are processed and sent in overlapping tasks to optimize the data processing time, with the value of sub-data pieces determined by network status and data processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the entire first data is sent to another communication node for processing and then the processing result is sent back, then the data processing can be completed, but the time consumed is excessive
Solution Approach 1:
The patent divides the entire first data into multiple pieces of sub-data before transmission. Each piece of sub-data is processed separately by different communication nodes, and the processing results are then aggregated. This segmentation approach parallelizes the processing workflow, allowing multiple operations to occur simultaneously rather than sequentially, thereby reducing overall processing time and improving data processing efficiency.
2Loss of time
If data is divided into r pieces of sub-data for parallel processing, then the processing time is reduced, but the system complexity increases
Solution Approach 1:
The patent dynamically determines the value of r (number of sub-data pieces) based on real-time network status and data processing load conditions. This dynamic adjustment allows the system to optimize the balance between processing speed and complexity - when resources are abundant, r can be increased for faster processing; when resources are constrained, r is reduced to maintain system stability. This dynamic parameter adjustment resolves the contradiction by making the system adaptable rather than static.
Solution Approach 2:
The system implements feedback mechanisms that continuously monitor network status and processing load, using this information to adjust the data division strategy. The feedback loop allows the system to learn from past performance and optimize future processing operations, balancing the trade-off between processing time reduction and system complexity management through data-driven decision-making.
Data Source
AI summary
A data processing method includes obtaining first data. The data processing method further includes determining a value of r based on a first parameter and/or a second parameter, where r indicates to divide the first data into r pieces of sub-data, the r pieces of sub-data are separately processed by a neural network layer in a machine learning model, a processing process of the first data is divided into tasks of sending the r pieces of sub-data, tasks of processing the r pieces of sub-data, and tasks of sending processing results of the r pieces of sub-data, and time occupied by tasks can overlap. The first parameter indicates a network status of a communication network used to send the sub-data, and the second parameter indicates a load status of a communication node configured to process the sub-data.