Direct-Connect Routing Schedules for Uneven Cluster Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for scheduling and routing in direct connect topologies, such as recursive halving, fail to efficiently utilize the varying bandwidths between intracluster and intercluster connections, leading to wasted bandwidth and prolonged data transmission times in computationally intensive applications.
Innovation Solution
A method and system that normalize intracluster and intercluster connections to a common form, allowing for simultaneous use of all connections to optimize routing and scheduling, and dynamically adjust routing schedules based on link performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If clusters operate in parallel and then share data using recursive halving, then simplicity of implementation is maintained, but bandwidth utilization becomes inefficient
Solution Approach 1:
The patent changes the scheduling parameters by introducing time steps and splitting coefficients to dynamically allocate bandwidth between intracluster and intercluster links. Instead of fixed sequential operations, the system uses parameterized scheduling that optimizes bandwidth utilization while maintaining implementation simplicity through automated controller management.
Solution Approach 2:
The system transitions from static sequential operations to dynamic scheduling where data transmission can occur simultaneously across different links at different time steps. The controller dynamically adjusts routing schedules based on workload constraints and bandwidth availability, enabling efficient parallel utilization of both intracluster and intercluster connections.
2Device complexity
If recursive halving is used for data sharing, then algorithm simplicity is preserved, but transmission time increases
Solution Approach 1:
The patent segments data transmission into multiple time steps and splits coefficients, allowing different portions of data to be transmitted simultaneously through different paths. This segmentation enables parallel processing of data across intracluster and intercluster links, reducing total transmission time while maintaining manageable algorithm complexity through structured decomposition.
Solution Approach 2:
The system adds a time dimension to the traditional spatial routing approach. Instead of only considering which links to use, the scheduler also determines when to use each link through time steps and splitting coefficients. This dimensional addition enables simultaneous utilization of multiple links, reducing transmission time without proportionally increasing algorithm complexity.
3Ease of operation
If intracluster and intercluster links are treated separately, then routing simplicity is maintained, but overall network efficiency decreases
Solution Approach 1:
The patent merges the scheduling of intracluster and intercluster links into a unified framework. The controller treats both link types together, using equivalency relationships and splitting coefficients to coordinate their operation. This merging enables simultaneous utilization of both link types while maintaining routing simplicity through automated coordination rather than separate manual scheduling.
Solution Approach 2:
The scheduling system achieves universality by handling both intracluster and intercluster links through the same mechanism. The equivalency-based approach and splitting coefficient framework apply uniformly to both link types, allowing a single scheduling algorithm to optimize overall network efficiency without requiring separate specialized routing logic for each link type.
Data Source
AI summary
Aspects of this disclosure relate to determining a first bandwidth of one or more intracluster links between one or more nodes of a first cluster; determining a second bandwidth of one or more intercluster links between the first cluster and one or more clusters; determining an equivalency between the one or more intracluster links and the one or more intercluster links based on the first bandwidth and the second bandwidth; and preparing a routing schedule based on the equivalency.