Network Switch Collective Operations in Distributed Architectures

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

As distributed computing systems grow in size and scale, they face challenges with bandwidth and hardware strain due to the inefficiencies in retrieving and processing data across multiple interconnected nodes, particularly in high-performance computing environments.

Innovation Solution

Implementing a network switch-based system that performs switch-based collective operations by identifying and prioritizing sub-operations, transmitting unicast 'get' messages to specific nodes, and iteratively calculating results, reducing network traffic and load on cache, memory, and processors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multicast techniques are used to retrieve distributed information across multiple computing nodes, then data retrieval capability is improved, but network bandwidth and hardware resources become strained

Engineering Contradiction:

Improvedata retrieval capabilityVSAvoidnetwork bandwidth and hardware resource strain

Core Design Contradiction:

ProductivityVSLoss of energy

Solution Approach 1:

The patent segments the collective operation into multiple sub-operations and distributes them across different computing nodes. Each node processes only its assigned sub-operations rather than all operations, reducing the computational burden and network traffic on individual nodes while maintaining overall system productivity.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a network switch as an intermediary device that performs collective operations. The switch acts as a mediator between computing nodes, receiving data from nodes and distributing processed results back to nodes, thereby reducing the direct communication overhead between nodes and minimizing network bandwidth strain.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Productivity

If distributed computing systems grow in size and scale, then compute performance and storage capacity are improved, but bandwidth and hardware availability become strained

Engineering Contradiction:

Improvecompute performance and storage capacityVSAvoidbandwidth and hardware availability management

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The network switch is designed to perform multiple functions: it acts as a standard network switch for data forwarding, a collective operation processor for computational tasks, and a data aggregation point for results. This multi-functionality allows the system to scale without proportionally increasing hardware complexity, as the switch handles multiple responsibilities that would otherwise require separate dedicated components.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS11044210B2Technologies for performing switch-based collective operations in distributed architectures

Publication Date: 2021.06.22 INTEL CORP

AI summary

Technologies for performing switch-based collective operations in a fabric architecture include a network switch communicatively coupled to a plurality of computing nodes. The network switch is configured to identify sub-operations of a collective operation of a collective operation request received from one of the computing nodes and identify a plurality of operands for each of the sub-operations. The network switch is additionally configured to request a value for each of the operands from a corresponding target computing node at which the respective value is stored, determine a result of the collective operation as a function of the requested operand values, and transmit the result to the requesting computing node. Other embodiments are described herein.