Distributed Gradient Reduction Algorithms for Mesh Topologies

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning systems face challenges in efficiently processing large datasets in parallel without a central parameter computation subsystem, leading to high latency and inefficiencies in gradient vector combination and parameter update operations.

Innovation Solution

The implementation of various algorithms such as single-path, pincer, rotated pincer, improved rotated pincer, iterative reduction, rotated iterative reduction, recursive reduction, and rotated recursive reduction algorithms for gradient reduction and broadcast operations across multiple processing units organized in circular, torus, and mesh topologies, enabling distributed and efficient processing of training data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If a central parameter computation subsystem is used, then gradient combination and parameter update operations are simplified, but system complexity and communication overhead increase

Engineering Contradiction:
Improvegradient combination operationVSAvoidsystem architecture
Core Design Contradiction:
Ease of operationVSDevice complexity

Solution Approach 1:

The patent extracts and eliminates the central parameter computation subsystem from the system architecture. Instead of having a central collector gather gradients from all processing units, the system uses distributed reduction operations where each processing unit participates in combining gradients locally, thereby simplifying the overall system structure while maintaining operational effectiveness.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The processing units perform gradient combination operations autonomously through distributed reduction algorithms. Each processing unit contributes to the reduction process without requiring coordination from a central authority, enabling self-organized gradient aggregation that eliminates the need for complex central management infrastructure.

Inventive Principle:
Principle #25Self-service

2Device complexity

If distributed processing without central subsystem is implemented, then system complexity is reduced, but communication latency and processing inefficiency increase

Engineering Contradiction:
Improvesystem architectureVSAvoidcommunication latency
Core Design Contradiction:
Device complexityVSLoss of time

Solution Approach 1:

The patent segments the gradient data into disjoint sets distributed across different processing units. Each processing unit handles a specific subset of gradient elements, and the reduction operation is divided into multiple stages where data is progressively aggregated. This segmentation enables parallel processing while controlling communication latency through structured data distribution.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a hierarchical dimension to the distributed reduction process. Instead of a flat single-stage reduction, the system organizes processing units in hierarchical levels with reduction operations proceeding through multiple stages. This dimensional organization enables more efficient communication patterns and reduces overall latency by processing gradients in parallel across different hierarchical levels.

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

3Ease of manufacture

If standard reduction algorithms are used, then implementation is simple, but processing speed and efficiency are insufficient for large datasets

Engineering Contradiction:
Improvealgorithm implementationVSAvoidtraining speed
Core Design Contradiction:
Ease of manufactureVSProductivity

Solution Approach 1:

The patent implements dynamic adaptation in the reduction algorithm based on the specific topology of the processing units. The reduction strategy changes depending on whether units are arranged in linear, tree, mesh, or other configurations, optimizing communication patterns for each topology. This dynamic approach maintains implementation simplicity while significantly improving processing speed for large datasets.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent performs preliminary organization of processing units into topological structures before executing reduction operations. By pre-establishing the hierarchical or network topology, the system can optimize communication paths and reduce the number of communication steps required during gradient aggregation, thereby improving training speed without complicating the reduction algorithm itself.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3859543B1Parallel processing of reduction and broadcast operations on large datasets of non-scalar data
Publication Date: 2024.08.21 GOOGLE LLC
  • EP3859543B1 patent drawingFigure 1~2
  • EP3859543B1 patent drawingFigure 3
  • EP3859543B1 patent drawingFigure 4A~4B

AI summary

Methods, systems, and apparatus, including instructions encoded on storage media, for performing reduction of gradient vectors and similarly structured data that are generated in parallel, for example, on nodes organized in a mesh or torus topology defined by connections in at least two dimensions between the nodes. The methods provide parallel computation and communication between nodes in the topology.