Data Center Traffic Demand Generation Using LDA Topic Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data center network traffic models are inadequate for simulating real-world scenarios due to assumptions of all-to-all spatial distribution and Poisson-distributed flow intervals, failing to accurately capture the heterogeneous and multidimensional characteristics of traffic demand in data center networks.
Innovation Solution
A method involving equal-frequency binning discretization and latent Dirichlet allocation (LDA) probability topic modeling to generate traffic demand data, which includes source and destination addresses, flow intervals, and flow sizes, creating high-dimensional feature and joint probability distribution matrices for realistic traffic simulation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional traffic models assume all-to-all spatial distribution and Poisson-distributed flow intervals, then the model complexity is low and easy to implement, but the model cannot accurately capture the heterogeneous and multidimensional characteristics of real traffic demand
Solution Approach 1:
The patent segments the traffic demand data by dividing the continuous flow interval and flow size attributes into discrete bins using equal-frequency binning. This segmentation transforms the complex continuous distribution into manageable discrete categories, enabling accurate characterization of heterogeneous traffic patterns while maintaining computational feasibility through structured data organization
Solution Approach 2:
The patent introduces latent topic dimensions through LDA modeling to capture the hidden multidimensional structure of traffic demand. By transforming the original four-dimensional traffic data (source address, destination address, flow interval, flow size) into a latent topic space, the model reveals underlying patterns and correlations that are not apparent in the original feature space, thereby accurately capturing heterogeneous characteristics
2Reliability
If traffic models use simplified assumptions for flow interval and flow size distribution, then the computational requirements are low, but the generated traffic data does not reflect real-world data center service scenarios
Solution Approach 1:
The patent performs preliminary equal-frequency binning discretization on flow interval and flow size data before applying LDA modeling. This preprocessing step organizes the data into structured bins in advance, reducing the computational complexity of subsequent LDA training while preserving the essential heterogeneous characteristics of traffic demand for realistic simulation
Solution Approach 2:
The patent transforms the continuous parameters of flow interval and flow size into discrete binned categories through equal-frequency binning. This parameter transformation reduces the computational burden of modeling continuous distributions while maintaining the ability to accurately represent the heterogeneous nature of real traffic patterns through the discrete category structures
Data Source
AI summary
A method for generating traffic demand data of a data center network includes: acquiring traffic demand samples each including a source address, a destination address, a flow interval, and a flow size; acquiring a first interval number by performing equal-frequency binning discretization processing according to the flow interval and acquiring a second interval number by performing equal-frequency binning discretization processing according to the flow size; determining a traffic demand subset according to the source address and the destination address, and acquiring a first parameter matrix and a second parameter matrix by training a latent Dirichlet allocation probability topic model according to the traffic demand subset; and generating the traffic demand data according to the first interval number, the second interval number, the first parameter matrix, and the second parameter matrix.