Distributed ML Training Job Server Bandwidth Bottleneck

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training machine learning models in large-scale distributed systems is computationally demanding and faces bottlenecks due to high communication bandwidth requirements, limiting training capacity and efficiency.

Innovation Solution

A method for training multiple machine learning models in a distributed computer system using a job server that allocates training jobs to groups of compute nodes, allowing these nodes to execute and update parameters locally, reducing the need for frequent communication with a parameter server and optimizing bandwidth usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If a single parameter server is used to synchronize parameters for the entire training task, then the model training can be coordinated centrally, but the communication bandwidth requirement becomes very significant and the parameter server becomes a bottleneck

Engineering Contradiction:
Improvecentralized parameter synchronizationVSAvoidcommunication bandwidth
Core Design Contradiction:
Extent of automationVSQuantity of substance

Solution Approach 1:

The system divides the distributed training task into multiple independent training groups, each with its own parameter server. Instead of one centralized parameter server handling all parameter synchronization, multiple parameter servers operate independently within their respective training groups, segmenting the communication load and eliminating the single-point bottleneck.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system transitions from a single-dimension centralized parameter server architecture to a multi-dimensional architecture where multiple parameter servers operate in parallel across different training groups. This dimensional expansion distributes the communication bandwidth requirement across multiple servers rather than concentrating it on one server.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If multiple machine learning models are trained simultaneously on multiple machines, then the training capacity increases, but the required communication bandwidth increases even more and the parameter server becomes a bottleneck

Engineering Contradiction:
Improvetraining capacityVSAvoidcommunication bandwidth
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The system segments the training task into multiple independent training groups, where each group trains a specific machine learning model with its own dedicated parameter server. This segmentation allows multiple models to be trained simultaneously without competing for the same communication bandwidth resources, as each training group operates independently with its own parameter synchronization channel.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Each parameter server is designed to be universal within its training group, handling parameter synchronization for the specific models assigned to that group. This multi-functional design allows the system to scale horizontally by adding more parameter servers and training groups, thereby increasing training capacity without proportionally increasing the communication burden on any single server.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Adaptability or versatility

If training involves large volume of data communication between parameter server and other machines, then the model can be trained on distributed system, but the required communication bandwidth becomes very significant

Engineering Contradiction:
Improvedistributed training capabilityVSAvoidcommunication bandwidth
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The system segments the large volume of parameter data into smaller, manageable units that are handled by multiple parameter servers operating in parallel. Each parameter server manages a subset of the total parameters, reducing the communication bandwidth requirement for each individual server while maintaining the overall distributed training capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system introduces multiple parameter servers as intermediaries between the compute nodes and the centralized control. Each parameter server acts as a local mediator for its training group, handling parameter synchronization locally rather than requiring all compute nodes to communicate directly with a single centralized server, thereby reducing the total communication bandwidth requirement.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3593247B1Training machine learning models on a large-scale distributed system using a job server
Publication Date: 2022.11.16 MIDEA GROUP CO LTD
  • EP3593247B1 patent drawingFigure 1
  • EP3593247B1 patent drawingFigure 2A~2C
  • EP3593247B1 patent drawingFigure 3

AI summary

A computer system for training machine learning models includes a job server and a plurality of compute nodes. The job server receives jobs for training machine learning models and allocates these training jobs to groups of one or more compute nodes. The allocation is based on the current requirements of the training jobs and the current status of the compute nodes. The training jobs include updating values for the parameters (e.g., weights and biases) of the machine learning models. Preferably, the compute nodes in the training group communicate the updated values of the parameters among themselves in order to complete the training job.