Http:// Huirong Fu and Edward W. Knightly Rice Networks Group Aggregation and Scalable QoS: A Performance Study.

http://www.ece.rice.edu/networks

Huirong Fu and Edward W. KnightlyRice Networks Group

Aggregation and Scalable QoS: A Performance Study

Edward W. Knightly

Problem: Scalability of Admission Control

Goal: provide predictable and controlled performance to Internet flows

Limitations of current approaches

– Intserv requires state communication and storage for each flow Scalability and deployability limitations

– Diffserv is simple and scalable but cannot quantify or control flow service quality (unless over-provisioned) Weaker service model

Edward W. Knightly

Can We Simultaneously Achieve ...?

High utilization

Scalability (not micro-managing flows)

Strong service model (e.g., suitable for VOIP)

– Internet (YYN)– Phone Network (NNY)– Intserv/ATM (YNY)– Diffserv (YYN)

Edward W. Knightly

IntServ with Aggregation

Ingress routers make “bulk” or aggregate core resv. – adjust as necessary

Core routers do not manage state, process signaling messages, and make reservations for every flow

A g g re g a te R e que st

F lo w R e que st

Edward W. Knightly

How Effective is Aggregation? … It Depends…

One extreme: traffic demand is relatively constant– Rarely signal core to adjust aggregate reservation– Achieve all three!

Other extreme: demand varies quickly and dramatically

(rapid and highly variable flow arrivals and departures)

1. True demand mismatches aggregate reservation Incorrectly block flows and under utilize network

2. Rapidly adjust aggregate reservation to track demand Lose signaling gain, default back to unscalable Intserv

Important role of timescales and variance of the traffic demand

Edward W. Knightly

Outline

Simple traffic and theoretical model to study aggregation

Validation and basic conclusions on timescales and variance

Remove assumptions of the basic model via simulations– Other primary demand functions– Correlation in secondary demand (multi-scale)

Trace driven simulations– Model validation– Insights into more realistic scenarios

Goal:devise framework to understand perf. of aggregation

Edward W. Knightly

Basic System Model

Aggregation system:– Ingress admits flow if sufficient bulk reservation

If new flow rate plus current demand < current agg. reservation

– Adjust bulk reservation level every seconds Assume perfect prediction for next seconds Well-defined control time scale

Assume bottleneck link C and N aggr.’s

Intserv admits flow if Ci i'

N 'N

j

11 '

1

e d g e ro u t e r

c o re ro u t e r

b o t t le n e c k lin k

flo wj

i

a n d lin k i

c la s s i

Cj

ii j

Edward W. Knightly

Simple Model of Aggregate Demand

Primary demand– Sinusoid with period

T, amplitude a, and random phase

Secondary demand– White noise– Uniform distribution U[-b,b]

+=

Demand time scale T

Demand variance also due to white noise

j

ii

a

b

T

Edward W. Knightly

Control and Demand Time Scales

IntServ Static Aggr. ResvDynamic Aggr. Resv

Recall: control time scale ; demand time scale T Intserv (=0) and static aggregate reservation (=T) upper

and lower bound performance Note: reserved resource utilization

Edward W. Knightly

Example Analytical Result

Overload probability - ratio of overloaded traffic (not admitted) to the total demand

Derived as:

N

iiii

T

k

T

k

N

iki

N

ol

baT

m

CfT

P N

i

1

1 1 1,

2

)(1

where

)2

cos(max1

, sT

amf iiksk

ki

ii

i

Edward W. Knightly

Overload and Control Time Scale

Performance continuum between Intserv and static reservation

If 0.01T, aggregation is near ideal

Given limit of signaling system, can determine achievable performance

Theoretical model tracks simulation results

T

Edward W. Knightly

Reserved Resource Utilization

RRU = fraction of reserved capacity utilized

Intserv is 1

Faster signaling better tracks demand, with

0.01T near perfect

T

Edward W. Knightly

Variance of the Secondary Demand

Demand variance degrades performance for– Ex. For perf. within 20% of Intserv’s, need var < .05, or

secondary demand range < .39 times primary range

16/T

T

var=0

var=0.01

var=0.33

+

0

0

Edward W. Knightly

Outline

Simple traffic and theoretical model to study aggregation

Validation and basic conclusions on timescales and variance

Remove assumptions of the basic model via simulations– Other primary demand functions– Correlation in secondary demand (multi-scale)

Trace driven simulations– Model validation– Insights into more realistic scenarios

Goal:devise framework to understand perf. of aggregation

Edward W. Knightly

Alternate Primary Demand Models

Different periodic functions with identical mean, variance, and period have little impact

Edward W. Knightly

Alternate Secondary Demand Models

Small impact, especially for smaller T2, smaller b

Uncorrelated

Correlated

T2=T/4

+

Edward W. Knightly

Trace-Driven Simulation Sources

Qbone trace (m 56.8 Mb/sec, var 191, T 24 hours, s 5 min) NLANR trace (m 0.74 Mb/sec, var 0.45, T 24 hours, s 1 sec) Caveat: all traffic vs. real-time flows

Edward W. Knightly

QBone Simulation and Model Predictions

System Variance is moderate

b/a=0.42

If =T/72, aggregate resv. achieves utilization of 97% of IntServ’s

Model Theoretical model retains

predictive capability

Primary + secondary outperforms primary only

Edward W. Knightly

NLANR Simulation and Model Predictions

System High variance in

secondary demand hinders performance (b/a=1.9)

If =0.01T, agg. achieves utilization of 44.2% of IntServ’s

ModelSecondary demand critical for model

Captures basic trend with larger prediction errors

Edward W. Knightly

Impact of Number of Aggregate Demands

Each aggregate introduces quantization error

Effect is cumulative and most visible for large

Could reverse trend via inter-aggregate statistical multiplexing or “merging”

Edward W. Knightly

Impact of Merging

Merge multiple aggregates into 1 vs. each independent

Significant performance improvements, especially when

Gains from statistical smoothing of multiplexed flow

T

Edward W. Knightly

Impact of Demand Phase

What if all aggregates are synchronized?

Performance degrades aggregation and IntServ

A capacity planning issue

)0(

Edward W. Knightly

Summary of Factors Affecting Aggregation

Major factors

– Demand time scale T, control time scale , variance

– < .01 T, and moderate variance is ideal

– Simple analytical model captures these effects

Minor factors

– Correlation structure of primary demand

– Existence of correlation (vs. white noise) in secondary demand

– Network topology (multiple bottlenecks)

Other Factors

– # of aggregates (-), merging (+), phase (- to all)

Edward W. Knightly

Conclusions

Proposed a simple model for aggregate traffic

Derived closed-form expressions for the system’s key performance metrics

Provide a methodology to determine the regime under which aggregation is an accurate and high-performance mechanism

http://www.ece.rice.edu/networks

Edward W. Knightly

Demand Model

Demand and Aggregation Model

]2,0[~ Ui

],,[~ iii bbUtZ 0s0)()( stZtEZ ii for

DemandSecondary

DemandPrimaryDemandAggregate

)()2

cos()( tZtT

amtr iiiii

Aggregate Demand, Request and Reservation

Edward W. Knightly

Demand Time Scale

To achieve performance within 10% IntServ, hours,

for minutes57.1T

9.5

Edward W. Knightly

NLANR (5 Minutes Average) Simulation and Model Predictions

Mean same, variance 0.45-->0.32

Since b/a decreases 1.9-->0.68, for 0.01T, aggregation performs better

Edward W. Knightly

”Sketch” Derivation ofOverload Probability

Consider aggr. resv. requests occur at identical epochs

Decouple the impact of primary and secondary demands

– Primary demand: odd symmetric

– Secondary demand: ADDITIONAL bandwidth must be reserved since

• Conditioning on the relative phases of different aggregates

T,,2,,0

1max0

ii

sbsZP

ib

Edward W. Knightly

Impact of Network Topology

Little impact– Large T incurs slight deviation according to the number of

contention points

1 5 5

3 1 11 5 5

1 5 5

1 5 5

1 5 5

1 5 5

1 5 5

1 5 51 5 5

1 5 5

1 5 5 M b p s1 5 5 M b p s

1 5 5 M b p s

1 5 5 M b p s

1 5 51 5 5 1 5 5

89.7

89.789

.7

89.7

1 5 5

3 1 1

1 5 5

8 9 . 7

8 9 . 7

8 9 . 7

8 9 . 7

8 9 . 7

8 9 . 7

8 9 . 7

8 9 . 7

6 2 2

1 5 5

1 5 53 1 1