Distributed Neural Network Training Inside RDBMS Clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current databases lack the capability to support deep neural networks at scale natively, leading to unnecessary data movement and limited scalability when training deep neural networks, especially with very large datasets.
Innovation Solution
A relational database management system (RDBMS) trains local neural network models in parallel using local data without moving data between compute units, and aggregates these models into a global neural network model, enabling training at scale within the database.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If data is moved from database to application for deep neural network training, then training can be performed using external frameworks, but data movement takes time, duplicates storage, and wastes compute cycles
Solution Approach 1:
The patent merges the database system with deep neural network training capabilities by integrating TensorFlow into the database engine. This allows training operations to be executed directly within the database without moving data externally, eliminating data transfer time and duplication while maintaining the ability to use sophisticated training frameworks.
Solution Approach 2:
The database system is enhanced to perform multiple functions: traditional relational processing plus deep neural network training. By embedding TensorFlow support directly in the database, the system can handle both data management and complex machine learning tasks within a single unified platform, eliminating the need for separate training infrastructure.
2Ease of manufacture
If centralized training is used, then training can be performed with existing frameworks, but scalability is limited when datasets do not fit in single machine storage
Solution Approach 1:
The patent segments the training process across multiple database nodes, each handling a portion of the dataset. By distributing data and computation across the database cluster, the system can train on datasets larger than single-machine capacity while maintaining the simplicity of centralized training coordination through the database management system.
Solution Approach 2:
The solution moves from single-machine training to multi-node distributed training by leveraging the database's cluster architecture. This dimensional expansion allows the system to handle larger datasets by utilizing the collective storage and compute resources across multiple nodes while maintaining a unified training interface.
3Adaptability or versatility
If separate infrastructures for big data management and deep neural network training are used, then existing frameworks can be applied, but unnecessary data movement and duplication occur
Solution Approach 1:
The patent combines big data management and deep neural network training into a single integrated infrastructure. By embedding TensorFlow within the database, the system eliminates the need for separate training infrastructure and associated data movement overhead, reducing energy consumption while maintaining framework compatibility.
Data Source
AI summary
An apparatus, method and computer program product for neural network training over very large distributed datasets, wherein a relational database management system (RDBMS) is executed in a computer system comprised of a plurality of compute units, and the RDBMS manages a relational database comprised of one or more tables storing data. One or more local neural network models are trained in the compute units using the data stored locally on the compute units. At least one global neural network model is generated in the compute units by aggregating the local neural network models after the local neural network models are trained.