Parallel Network Switch Routing for Communication Contention

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing parallel computer systems face performance issues due to communication path contention during collective communication, particularly in all-to-all communication, as existing techniques are not adaptable to various network topologies.

Innovation Solution

A parallel computer system with network switches connected to multiple nodes, where each node determines a destination node to avoid communication path contention by using systematic and static routing, ensuring that data is transmitted through different switches, thereby preventing concentration on specific paths.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If data is transmitted through network switches in a parallel computer system, then communication can be established between nodes, but communication path contention occurs causing performance degradation

Engineering Contradiction:
Improvecommunication speedVSAvoidcommunication path contention
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The invention segments the communication paths by dividing the network into multiple virtual channels or logical paths within the network switches. This allows multiple data transmissions to occur simultaneously through different segments of the network infrastructure, preventing contention on any single physical path while maintaining full network connectivity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention introduces an additional dimensional layer of routing by implementing multi-dimensional network switch configurations where data can be routed through different spatial and logical dimensions. This allows communication paths to be differentiated not just by physical connections but by virtual channel assignments, enabling parallel transmissions without interference.

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

2Adaptability or versatility

If existing communication algorithms are applied, then communication can be performed, but they are not adaptable to various network topologies

Engineering Contradiction:
Improvetopology adaptabilityVSAvoidalgorithm complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The invention creates a universal communication algorithm that functions across multiple network topologies by implementing topology-agnostic routing logic. The network switches and nodes can dynamically adapt to different topological configurations (mesh, torus, hypercube, etc.) using the same fundamental algorithmic framework, eliminating the need for topology-specific implementations while maintaining optimal performance.

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

Solution Approach 2:

The invention utilizes parameter changes by allowing the communication algorithm to dynamically adjust routing parameters, message dimensions, and switch configuration settings based on the detected network topology. This enables a single algorithm to efficiently operate across diverse topologies by modifying its operational parameters rather than requiring different algorithms for each topology type.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS10361886B2Apparatus and method for collective communication in a parallel computer system
Publication Date: 2019.07.23 FUJITSU LTD
  • US10361886B2 patent drawing
  • US10361886B2 patent drawing
  • US10361886B2 patent drawing

AI summary

A parallel computer system includes a plurality of network switches that are all connected to each other, and a plurality of nodes each connected to one of the plurality of network switches, where each network switch is connected to two or more nodes of the plurality of nodes. Each node determines a first destination node of data to be transmitted by the each node at a given time so that a first network switch connected to the first destination node is different from a second network switch connected to a second destination node of data transmitted by any node, other than the each node, which is connected to a network switch to which the each node is connected, and transmits data to the determined first destination node.