Adaptive Multi-Path Routing With Bin Reassignment for Congestion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for routing traffic in high-performance computing scenarios, such as large-scale or multi-tenant machine learning applications, result in inefficient and unstable routing due to artificial groupings of unrelated traffic, leading to congestion and excessive power consumption, as they rely on hash-based mechanisms that do not optimize the route of traffic.
Innovation Solution
Implementing a system that assigns packets to bins using a non-deterministic process, such as random or pseudo-random selection, and updates bin associations in response to adaptive routing notifications (ARNs) to achieve a steady state, ensuring even distribution and load balancing across multiple paths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If hash-based mechanisms are used to route traffic, then routing decisions can be made deterministically, but traffic distribution becomes unstable and congestion occurs
Solution Approach 1:
The patent applies dynamics by transitioning from static hash-based bin assignment to dynamic bin association that adapts to network conditions. Bins are dynamically reassigned to different egress ports based on real-time congestion feedback, allowing the system to maintain determinism while adapting to changing traffic patterns and avoiding congestion hotspots.
Solution Approach 2:
The patent implements feedback mechanisms through adaptive routing notifications (ARNs) that provide congestion information from egress ports back to ingress ports. This feedback loop enables the system to adjust bin assignments dynamically, resolving the contradiction between deterministic routing and stable traffic distribution by using congestion feedback to guide routing decisions.
2Productivity
If multiple paths are used for load balancing, then network throughput increases, but power consumption increases unnecessarily during low traffic periods
Solution Approach 1:
The patent applies partial action by activating only the necessary number of network paths based on actual traffic demands. During low traffic periods, fewer paths are active, reducing power consumption. During high traffic periods, more paths are activated to maintain throughput, optimizing the balance between productivity and energy consumption.
Solution Approach 2:
The system dynamically adjusts the number of active paths based on traffic conditions and congestion feedback. The adaptive routing mechanism enables the switch to scale path utilization according to demand, avoiding the excessive power consumption of maintaining all paths active during low traffic while preserving high throughput capability when needed.
3Device complexity
If conventional hash-based routing is used, then routing implementation is simple, but traffic distribution is inefficient and causes congestion
Solution Approach 1:
The patent introduces adaptive routing notifications (ARNs) as an intermediary mechanism that carries congestion information from egress ports to ingress ports. This intermediary enables efficient traffic distribution by providing real-time feedback about path conditions, allowing the system to optimize routing without requiring complex centralized control or sacrificing implementation simplicity.
Solution Approach 2:
The feedback mechanism through ARNs enables the system to achieve efficient traffic distribution while maintaining relatively simple implementation. The congestion feedback allows local adaptive routing decisions that improve productivity without requiring complex global optimization algorithms, balancing implementation complexity with traffic distribution efficiency.
Data Source
AI summary
A device, communication system, and method are provided. In one example, a system for routing traffic is described that includes circuits to receive a plurality of packets associated with a destination, determine a plurality of bins based on the destination of the packets, wherein each bin of the plurality of bins is associated with a respective port of the plurality of ports, and, for each packet of the plurality of packets, select a bin of the plurality of bins and forward the packet via the port associated with the selected bin. Described circuits also receive an adaptive routing notification packet via a first port associated with a first bin and in response associate the first bin with a second port and de-associate the first bin with the first port.