Supercomputer Network Topology Using Grouped Global Links
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing supercomputer network topologies face limitations in scalability and efficiency due to hardware constraints, particularly in connecting multiple servers through networks, necessitating improved methods for forming efficient network topologies.
Innovation Solution
A device and method for generating a network topology by selecting network groups based on the number of global links and switch ports, forming global links between these groups to create a topology that includes a large number of 1-extra-hop routes, and using adaptive routing to manage traffic flow.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If more servers are connected through network to improve supercomputer performance via scale-out, then the performance of supercomputer is improved, but the network topology complexity and difficulty of managing connections increases
Solution Approach 1:
The network is segmented into multiple network groups (NG0-NG7) with hierarchical structure. Each network group contains switches and servers organized in manageable units. This segmentation allows the large-scale network to be divided into smaller, more manageable components while maintaining overall connectivity and performance.
Solution Approach 2:
The patent introduces a hierarchical dimension to the network topology by organizing switches and servers into multiple levels and network groups. This multi-dimensional structure (with dimensions for network groups, switches, and servers) simplifies the management of connections by providing structured layers of organization rather than flat pairwise connections.
2Ease of manufacture
If traditional network topologies are used to connect servers, then implementation is straightforward, but scalability is limited due to hardware constraints
Solution Approach 1:
The network topology is designed to be dynamic and adaptive rather than static. The system can dynamically select paths between network groups based on current conditions, and the hierarchical structure allows for flexible expansion. Network groups can be added or removed without requiring complete reconfiguration of the entire network, enabling scalable growth while maintaining ease of implementation through standardized interfaces.
3Productivity
If more global links are added to connect network groups, then network connectivity and performance improve, but hardware constraints and cost increase
Solution Approach 1:
Instead of adding more direct links between all pairs of network groups (which would require O(n²) links), the patent introduces a hierarchical dimension with intermediate network groups. This reduces the number of global links required from quadratic to linear complexity, as links are shared across multiple network groups through the hierarchical structure.
Solution Approach 2:
The hierarchical network group structure can be copied and replicated to expand the network. Standardized network groups with defined interfaces can be instantiated multiple times, allowing the network to scale by copying proven working units rather than designing custom connections for each expansion scenario.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A device includes: a processor configured to execute instructions in a memory, wherein, the instructions are configured to cause the processor to: select a first network group from among network groups, wherein the first network group is selected from among the network groups based on numbers of global links connecting the network groups, respectively; obtain a first list of network groups, among the network groups, that are not connected to the first network group; select a second network group and a third network group from among the network groups in the first list based on connections between the network groups in the first list; and based on the selecting of the first, second, and third network groups, form a topology including global links connecting the first network group, the second network group, and the third network group each to each other.