Rail-Based Link Status Propagation for Large Network Fabrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In network fabrics with spine-leaf architecture, frequent communication of link status information consumes significant bandwidth and processing power, adversely affecting throughput, latency, and workload resumption, particularly in large GPU clusters.
Innovation Solution
Implement a rail-based network architecture where link status information is propagated on a rail basis, reducing data transmission by each leaf switch, and aggregated at host devices, using LLDP messages with OUI Type-Length-Value fields and sequence identifiers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If link status information is communicated frequently across all leaf switches and endpoint servers, then network reliability and topology awareness are improved, but bandwidth consumption and processing power requirements increase significantly
Solution Approach 1:
The patent segments link status information propagation by rail identifier, dividing the large network into smaller manageable groups. Each leaf switch only propagates link status information for its associated rail identifier to matching rail identifiers at host devices, rather than broadcasting to all devices. This segmentation reduces the scope of communication while maintaining reliability for relevant network segments.
Solution Approach 2:
The patent implements local quality by making link status propagation selective and targeted. Each network device filters and processes only the link status information relevant to its rail identifier, rather than processing all link status information in the network. This localizes the processing burden and reduces overall network traffic while maintaining necessary reliability information.
2Loss of information
If link status information is transmitted for all 32,768 ports, then complete network topology awareness is achieved, but data transmission volume increases to 4,096 bytes per leaf switch
Solution Approach 1:
The patent segments topology information by rail identifier, so that each leaf switch only transmits link status information for ports associated with its rail identifier. This reduces the data transmission volume from 4,096 bytes per leaf switch to a much smaller amount, while still providing complete topology awareness through the aggregation of rail-specific information at host devices.
Solution Approach 2:
The patent introduces a new organizational dimension by grouping ports and switches by rail identifier. This dimensional reorganization allows the system to represent topology information more efficiently, reducing the total data volume needed while maintaining complete topology awareness through the structured aggregation of rail-specific information.
3Adaptability or versatility
If frequent communication of link status information is implemented, then dynamic adaptation to network changes is improved, but throughput and latency are adversely affected
Solution Approach 1:
The patent segments link status communication by rail identifier, reducing the frequency and volume of communications required for dynamic adaptation. Each leaf switch only communicates with host devices having matching rail identifiers, reducing the overhead associated with frequent updates while maintaining the ability to adapt dynamically to network changes in relevant segments.
Solution Approach 2:
The patent applies local quality by making each network device process and act on only the link status information relevant to its rail identifier. This localized processing reduces the computational overhead and communication frequency needed for dynamic adaptation, thereby improving throughput while maintaining adaptability to local network changes.
4Loss of information
If link status information is propagated to all host devices, then comprehensive network monitoring is achieved, but processing power consumption increases
Solution Approach 1:
The patent segments link status information processing by rail identifier, so that each host device only processes link status information relevant to its rail identifier. This segmentation dramatically reduces processing power consumption while maintaining comprehensive network monitoring capability through the systematic aggregation of rail-specific information.
Solution Approach 2:
The patent implements local quality by having each host device process only the link status information relevant to its rail identifier rather than processing all link status information in the network. This localized processing reduces overall processing power consumption while maintaining comprehensive monitoring through the structured collection of rail-specific data.
Data Source
AI summary
Devices, systems, methods, and processes for link status propagation are provided. In modern networks, transmitting a single bit of information about link status of each host-side link in the network may require a large amount of data, consuming considerable bandwidth and processing power. To address these concerns, a network device having a plurality of host-side ports coupled to same ordinal processing units in a set of host devices and associated with a rail identifier is provided. The network device determines first link status information associated with communication links of the network device and receives second link status information from other network devices having rail identifiers that match the rail identifier of the network device. The network device transmits the first and second link status information to the same ordinal processing units and a host device aggregates link status information received at corresponding processing units to obtain a cluster wide view.