Optical Server Interconnection for Lower-Power Data Transmission
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies fail to provide a solution for full interconnection between servers in an optical switching network, leading to increased power consumption and inefficiencies in data transmission during distributed training of large models.
Innovation Solution
A communication method that determines a minimum quantity of computing devices required for each server to enable full interconnection, using a connection policy to connect servers in an optical networking architecture, and considers connection topology and aggregation manners to reduce overheads.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the capacity of the electrical switch is expanded to transmit data generated during training of the large model, then the data transmission capacity is improved, but the power consumption in the entire data transmission process is significantly increased
Solution Approach 1:
The patent replaces the electrical switching network with an optical switching network to transmit data. Optical switching uses light instead of electrical signals, which enables higher data transmission capacity without the power consumption increase associated with expanding electrical switch capacity. The optical switch connects NPUs across servers using optical fibers, providing a more energy-efficient solution for high-speed data transmission during large model training.
2Use of energy by moving object
If an optical switching network is used to transmit data generated during the training of the large model, then the power consumption is reduced, but a solution for connection between NPUs is not provided in conventional technology
Solution Approach 1:
The patent segments the connection problem into two parts: first determining the minimum quantity of computing devices required by each server to implement full interconnection, then establishing the connection policy. This segmentation allows the system to calculate the optimal number of computing devices needed and then configure the optical switching network accordingly, providing a complete solution for NPUs connection in the optical network.
Solution Approach 2:
The patent performs preliminary calculations to determine the minimum quantity of computing devices required before establishing the optical connections. By pre-calculating the required number of computing devices based on the quantity of servers, the system can then proceed to configure the connection policy and establish the optical network connections, ensuring that the network is properly designed before implementation.
3Productivity
If full interconnection between all servers is implemented in the optical switching network, then the data transmission efficiency is improved, but the complexity of determining connection policies increases
Solution Approach 1:
The patent changes the approach from considering all possible connections to focusing on the minimum quantity of computing devices required. By parameterizing the solution around the minimum number of devices needed for full interconnection, the system simplifies the connection policy determination while still achieving full interconnection between servers, thus reducing the complexity of managing the optical network.
Data Source
AI summary
This application relates to a communication method and apparatus, to implement a full interconnection between servers by connecting a computing device in a server included in a computing cluster to an optical switch. In this application, a management device obtains a quantity M of servers in the computing cluster and a quantity N of computing devices in each server; determines, based on the quantity M of the servers, a minimum quantity n of computing devices required by each server to implement a full interconnection between the M servers, where M, N, and n are positive integers; when n is less than or equal to N, determines a connection policy between the servers when the M servers are fully interconnected; and connects the M servers according to the determined connection policy to enable the M servers to be fully interconnected.