Traffic Driven Analysis of Cellular Data Networks Samir R. Das - - PowerPoint PPT Presentation

Traffic Driven Analysis of Cellular Data Networks

Samir R. Das Computer Science Department Stony Brook University

Joint work with Utpal Paul, Luis Ortiz (Stony Brook U), Milind Buddhikot, Anand Prabhu Subramanian (Alcatel‐Lucent Bell Labs)

Mobile Data Usage

Relatively little research on nature of mobile data traffic.

2

0.6 EB / month 10.8 EB / month

Forecast of Global Mobile Data Traffic

Source: CISCO VNI Mobile

1 Exabyte = 1 million Terabyte

Higher than the traffic volume in the entire Global Internet In 2006

3

Traffic Management Modeling and Forecasting Traffic Analysis

Flow Monitoring Tool SQL Database

Packet Flows

Internet

Flow Records

Mobility and Session Manager Radio Access Network

Measurement Infrastructure

Sample Results from Traffic Analysis

• Data collected from a nationwide 2G/3G network

circa 2007

– About 10K BSes, 1M subscribers.

• Significant traffic imbalance per subscriber and

per BS

– 1% of subscribers create more than 60% of load. – 10% of BSes experience more than 50% of load.

• Mobility is generally low

– More than 50% subscribers stick to just one BS daily. – Median radius of gyration is ~1 mile.

Sample Results from Traffic Analysis

• Mobility is predictable

– Subscribers are almost always found in their top 2‐3 most visited locations. – They return to the same location at the same time of the day with high probability.

• More mobile subscribers tend to generate more

traffic.

• Radio resource usage efficiency is very poor

– Much poorer for light users relative to heavy users.

Functional Influence Among BSes

• Model BS load as time series. Explore causal

relationships between pairs of time series.

• Granger Causality – Determines whether one time

series is useful in forecasting another when using an autoregressive model.

– Has been used in economics and neuroscience.

• Statistically significant causality exists among

neighboring BSes (roughly among half of the neighbors).

• Causality graph and causal path – Make a graph out of
• causality. Long paths exist in this graph (median = 15

hops, 90‐percentile = 37 hops).

Modeling Study

• Model BS traffic loads exploiting any

interactions/dependencies

– Exploit tools from machine learning. – Many possible directions – purely static/spatial, dynamic/temporal.

• Goals:

– Intellectual – broad understanding of any underlying structure would help future network architectures. – Utilitarian – models can help estimation/forecasting. Useful for various resource management.

• Assume load on n base stations are multi‐variate

Gaussian:

• Learn the parameters given a set of training data,

specifically the “inverse covariance matrix” , given a set of training data (p observations).

• is easier to estimate than and exposes

interesting properties.

Spatial Modeling Approach: Probabilistic Graphical Modeling

Mean vector Covariance matrix

1



Inverse Covariance Matrix: Properties

• If then load

variables and are conditionally independent, given the rest of the variables.

• Most problems produce a

`sparse’ model.

• Related to probabilistic

graphical models (e.g., Gaussian Markov Random Field).

1 2 3 4 5

Undirected Graphical Model ‐> Edge ‐> no edge Graph properties translate to probabilistic (in)dependencies

Xi

X j

Inference Problem

• Estimate load for BS i given

the load of a subset of BSes S as the conditional mean:

• Broad questions:

– How large should be S? Effort

• vs. accuracy tradeoff.

– How to choose S?

1 2 3 4 5 Measure only a subset and estimate the rest.

First Solve the Learning Problem

• Learn the inverse covariance matrix from

training data.

• How? Exploit relationship with linear

regression modeling.

– Express load of BS i as a linear function of all other BS loads and then regress: – Regression coefficients can be shown to be directly related to inv. cov. matrix elements.

Yi  X j j

i



 j

i

Sparse Models

• Sparse model ‐> many regression coeffs

are zero.

• Reduces danger of over‐fitting (lowering

variance). Also, computationally efficient.

• Introduce a regularization term in regression.

We used “Lasso” .

Empirical error Regularization term modeling penalty

 j

i

Regularization

• Cross‐validate using additional training

samples (not used for model creation).

• Use various values of to create different

models.

• Choose the one with max likelihood.



Data Processing

• Hourly load of 400 BSes covering 75 x 84 miles area.

Includes a busy downtown and surrounding suburbs.

• No temporal dimension in model. Create different

models for for different parts of the day (every 4 hours).

• Account for diurnal variation of load. Use residuals

from a fitting function. Residuals pass normality test.

Average Edge Length in the Model Graph

• Apparent spatial/regional significance.

In miles In hops (in Voronoi Graph)

Choosing the Measured Set S

• Greedy strategy – each iteration picks the BS that

minimizes the error estimate.

• Higher load first achieves almost similar performance.

Impact of Estimation Accuracy on Applications

• We understand the measurement complexity

(size of S) vs. Error tradeoff.

• But how much accuracy do we need? Need to

turn to applications

• Studied two applications

– Energy Management – Opportunistic Traffic Scheduling

Opportunistic Traffic Scheduling

• Similar to Smart Electric Grid – move non‐urgent traffic

from peak to off‐peak periods.

– What is non‐urgent? p2p, large downloads, sync, push, etc. – Who decides? User agent on mobile. May have multiple levels of priority or have deadlines to aid scheduling. – Carriers can incentivize such scheduling.

• Similar to QoS scheduling – but at a higher layer and at

a longer time scale.

• Two components in System Architecture

– Server (Scheduler) in core network. – User agent on mobile that coordinates with Server.

Time Line 2PM 2:30PM 3PM 3:30PM

Creates low-priority flow Deadline=2hr

20

Server (scheduler) in the core network

Solving the Scheduling Problem

• Several approaches possible based on how

flows are prioritized.

• But for any approach, server needs to be able

estimate current/future loads at all BSes.

– Also, needs to model/estimate subscriber mobility (separate problem).

• Poor estimation leads to poor scheduler

performance.

Evaluation Approach

• Trace‐driven simulator based on a capacity model
• f BSes.
• Opportunistic scheduling is meant to admit more

traffic but with the same network capacity.

• We use the same traffic trace always, but reduce

network capacity to demonstrate impact.

• Impact?

– Do low priority flows still finish within a reasonable time? – Are high priority flows impacted?

Results

• Low priority flows =

random subset of long‐ lived flows (over 25 mins), about 8% of all flows. Randomly chosen deadlines 1 ‐ 4 hours.

• Rest high priority.
• Scheduling epoch hourly.
• Only a subset of 400 BSes

are measured, rest estimated.

Conclusions

• Discovering structures in mobile traffic is a

rich area of study.

• Applications in network and resource

management.

25

Traffic Management Modeling and Forecasting Traffic Analysis

Questions?