Parallel Computing Node Groups for Overlapping Graph Reasoning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning frameworks and graph reasoning methods utilize computing resources inefficiently, leading to underutilization of CPU and computing card resources due to single-threaded operations and asynchronous computations.
Innovation Solution
A method for parallel processing of data that identifies logical relationships among computing nodes, forming parallel node groups for simultaneous computation, ensuring CPU and computing card resources are fully utilized by overlapping operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If single-threaded asynchronous computations are used for graph reasoning, then implementation simplicity is maintained, but CPU and computing card resources cannot be fully utilized
Solution Approach 1:
The patent segments the graph reasoning computation into multiple independent thread groups, where each thread group processes a subset of computing nodes. This segmentation allows parallel execution across multiple threads while maintaining manageable complexity through modular organization of computation tasks.
Solution Approach 2:
The patent transitions from single-threaded sequential processing to multi-threaded parallel processing by adding the thread dimension. Multiple thread groups operate simultaneously on different subsets of computing nodes, transforming the computation from a one-dimensional sequential process to a multi-dimensional parallel process, thereby fully utilizing CPU and computing card resources.
2Productivity
If data flow architecture is used for deep learning operators, then computational efficiency is improved, but adaptability to non-deep learning operations is reduced
Solution Approach 1:
The patent creates a universal computing node processing framework that can handle both deep learning operators and general graph reasoning operations. The unified data structure and processing pipeline accommodate different operation types, allowing the high-efficiency data flow architecture to serve multiple purposes beyond just deep learning.
Solution Approach 2:
The patent uses parameter-based configuration to adapt the data flow architecture to different operation types. By changing parameters such as computing node attributes, edge relationships, and operation specifications, the same architectural framework efficiently processes both deep learning operators and general graph reasoning tasks without sacrificing computational efficiency.
Data Source
AI summary
Provided are a method for the parallel processing of data, a device, and a storage medium. The method includes: identifying, from multiple first computing nodes, at least three first computing nodes which have a logical relationship are identified and defining the at least three first computing nodes which have the logical relationship as a first parallel node group, where the first parallel node group includes a first preceding node and at least two first subsequent nodes; acquiring a first input data model of the first preceding node and generating a first input tensor of the first preceding node; computing a first output tensor of the first preceding node according to the first input data model and the first input tensor; and acquiring a second input data model of the at least two first subsequent nodes and using the first output tensor as a second input tensor.