Incorporating Intra-flow Dependencies and Inter-flow Correlations for - - PowerPoint PPT Presentation

Incorporating Intra-flow Dependencies and Inter-flow Correlations for Traffic Matrix Prediction

Presenter: Kaihui Gao Tsinghua University

Catalog

• 1. Background
• 2. Related Works
• 3. Motivation
• 4. Data Analysis of Traffic Matrix
• 5. Challenge
• 6. Solution
• 7. Evaluation

Background

• Accurate TM prediction plays an important role in the areas
• f traffic engineering (TE), capacity planning, resource

allocation, congestion mitigation, and failure/anomaly detection in computer networks.

• For network traffic, as a kind of spatial-temporal series, its

forecasting has to face similar key challenges of modeling sequential information as the traditional time series forecasting problem. Furthermore, accurate forecasting requires addressing the challenges of modeling the complex and underlying spatial correlation.

Problem Formulation

• TM: Traffic Demand Matrix
• (TMt+1 ,TMt+2,…, TMt+q) = f(TMt,TMt-1,…,TMt-w+1)

1 2 3 4 5 6 7 8 9 10 11 12

Related Works

• ARMA, ARIMA (Auto-regressive Integrated Moving Average)
• linear structure cannot model the nonlinear patterns
• SVR (Support Vector Regression)
• significant effort is needed to tune its parameters
• Neural Networks (NNs) based approaches
• RNN: LSTM, GRU

State-of-the-art: RNN

• TMt-w+1

TMt-w+2 TMt-1 TMt … TMt+1

Motivation

• server

client1 client2 request request response response

• Popular service ——> positive correlations
• Bandwidth contention ——> negative correlations

Spatial correlation analysis

• Pearson Correlation Coefficient:

Observation 1: There exist abundant high correlations between flows in real WAN.

Spatial correlation analysis

• Observation 2: Most strongly-correlated flow pairs share

the same source or destination.

How to capture spatial correlations

• Explicitly
• high computational overhead
• curse of dimension
• time-varying
• Implicitly
• learning based
• convolutional neural network

Solution

• We propose a novel deep learning framework combined

CNN and RNN to predicting future traffic matrix.

• Combined of
• CNN capture the spatial correlations of every flow pair.
• RNN model the temporal variation of spatial correlations

and traffic.

Overview

Evaluation

• Datasets: Abilene and GEANT
• training set : test set = 8 : 2
• Performance Metric:
• Mean Square Error
• Mean Absolute Error
• Baseline: ARIMA, SVR, LSTM, CNN

Predicting Next Single TM

Predicting Next Multiple TMs

Traffic Engineering based on TM Prediction

• TE problem: given a network with edge capacities and a set
• f traffic demands of source-destination pairs, we try to find

an assignment of demands to paths (an ordered sequence of edges) that optimizes some objectives. A widely used TE

• bjective is minimizing the maximum link utilization (MLU).
• Multicommodity Flow (MCF) problem
• Baseline
• MCF based on real TMs
• MCF based on predicated TMs by CRNN
• MCF based on predicated TMs by LSTM
• shortest path routing (SP)
• Oblivious routing (OR)

Performance comparison

!