Flowlet Load Balancing via Local Weight Tables
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional load balancing methods in data center networks are complex and inefficient, leading to traffic congestion due to the need for real-time load status updates across all paths, which results in unsatisfactory load balancing effects.
Innovation Solution
A load balancing method that calculates weight values for equal-cost paths based on network topology, port status, and link bandwidth, allowing switches to perform flowlet-based load balancing using a locally stored equal-cost path weight table, reducing the need for real-time load status updates and simplifying the algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-time load status updates are performed across all paths, then load balancing accuracy is improved, but system complexity and processing overhead increase
Solution Approach 1:
The patent segments the load balancing decision-making process into two parts: (1) The controller performs global path load status detection and calculates weight values for equal-cost paths; (2) Switches locally store these weight values in equal-cost path weight tables and use them for flowlet-based load balancing decisions. This segmentation allows accurate load balancing while reducing the complexity at individual switches.
Solution Approach 2:
The controller performs preliminary actions by pre-calculating the weight values of equal-cost paths based on global network topology and link bandwidth information, then delivers these weight values to switches in advance. This allows switches to make load balancing decisions using pre-computed weights without performing complex real-time calculations, thereby improving accuracy while reducing operational complexity.
2Productivity
If global path load information is collected and processed, then load balancing effectiveness is improved, but processing time and algorithm complexity increase
Solution Approach 1:
The controller pre-calculates weight values for equal-cost paths based on global network information including topology and link bandwidth, then delivers these weights to switches in advance. This preliminary calculation eliminates the need for switches to perform time-consuming real-time global path analysis, achieving effective load balancing while reducing processing time.
Solution Approach 2:
Instead of having switches perform complex global load analysis, the controller creates a simplified representation (weight values) of the global network state and copies this information to switches via equal-cost path weight tables. This copying approach allows switches to make effective load balancing decisions using simple local lookups rather than complex real-time calculations.
3Measurement precision
If complex load balancing algorithms are implemented, then traffic distribution accuracy is improved, but computational overhead increases
Solution Approach 1:
The patent divides the computational workload: the controller performs the complex global optimization calculations to determine optimal weight values for equal-cost paths, while switches perform simple local lookups in their equal-cost path weight tables. This segmentation achieves accurate traffic distribution while minimizing computational overhead at switches.
Solution Approach 2:
The controller creates a simplified copy of the optimal load balancing strategy in the form of weight values stored in equal-cost path weight tables at switches. This copying approach allows switches to achieve accurate traffic distribution by referencing pre-computed weights rather than performing complex real-time calculations, thereby reducing computational overhead.
Data Source
Figure 1A~1B
Figure 2
Figure 3
AI summary
This application discloses a load balancing method and apparatus, and a device, and relates to the field of network technologies. The method includes: after a packet sent by a server is received, determining whether the packet is a first packet of a flowlet; if the packet is the first packet of the flowlet, determining a destination switch based on a destination address of the packet; determining a weight value of at least one equal-cost path associated with the destination switch in a stored equal-cost path weight table, where the equal-cost path weight table stores a correspondence between the at least one equal-cost path and the weight value; and scheduling, based on the weight value of the at least one equal-cost path, the packet onto a corresponding equal-cost path for transmission. In this application, after receiving the packet sent by the server, a switch detects the flowlet, and directly performs flowlet-based load balancing based on the locally stored equal-cost path weight table, instead of obtaining load statuses of all paths in real time. This greatly reduces complexity of a load balancing algorithm, and achieves a great effect.