Adaptive Resource Allocation: Self-Sizing for Next Generation Networks Michael Devetsikiotis Electrical & Computer Engineering North Carolina State University.

Adaptive Resource Allocation: Adaptive Resource Allocation: Self-Sizing for Next Generation NetworksSelf-Sizing for Next Generation Networks

Michael DevetsikiotisMichael DevetsikiotisElectrical & Computer Engineering

North Carolina State University

Dr. Qi Hao, Nortel Networks, Ottawa

Dr. Sandra Tartarelli, NEC Research, Germany

Dr. Matthias Falkner, Cisco, Germany

Dr. Jiangbin Yang, Lantern Communications

Fatih Haciomeroglu (M. Sc., graduated)

Srikant Nalatwad (Ph. D. candidate)

Peng Xu (Ph. D. candidate)

Robert Callaway (M. Sc. candidate)

2

Overview of InterestsOverview of Interests Measurement-based, adaptive resource allocation

use traffic measurements to improve congestion control include prices, QoS use statistical methods to predict, model and simulate efficiently

Open Loop: “Self-sizing” of ATM/MPLS via adaptation Preemption and re-routing methods

Closed Loop: Predictive Active Queue Management Predictive methods for Explicit Congestion Notification Games for loss networks (optical, wireless) with incomplete info

3

Self-Sizing Self-Sizing Using MeasurementsUsing Measurements Bandwidth allocation: make adaptive rather than static

Use bandwidth more efficiently while satisfying QoS constraints. Adaptive algorithms based on traffic measurements: large gains.

Plan: Separate into “data” and “control” parts Research and compare effective bandwidth techniques. Show efficiency gains, QoS delivery. Investigate measurement time scales and parameter settings. Research “globality” of information for larger networks (scaling). Study of gain vs. adaptation time scale.

Try out in simulation, then emulation C++, OPNET RTFM in open source (IETF), Linux test-bed

4

Simplified IllustrationSimplified Illustration

5

t,0X

tEelog

t

11lim

bQPb

1lim

b

Traffic Descriptor: Effective BandwidthTraffic Descriptor: Effective Bandwidth

The EB is a measure of the amount of bandwidth a source requiresto meet its QoS. From Large Deviation Theory:

•P(Q>b): required QoS•b: determined by max delay

cNK

1jjj

bebQP

Under the hypothesis of large buffers:

b

bQPlog

)( required bandwidth

For each traffic class within each SD pair:

C

)(N 11

)(N KK

• Theoretical Framework

6

Measured Effective BandwidthsMeasured Effective Bandwidths

S D

NAP

measurement point

Network

• For each traffic class within each SD pair the aggregated traffic is measured

t/N

1i

iX~

eN

tlog

t

1t,ˆ nn1nn BAˆˆ

nt,1t)1n(Xn e

Nt

logt

1A

ntN,1t)1n(NXn e

Nt

logt

1B

On-Line Estimation

Measurement window t

(n-1)

n

7

Optimization Model ExampleOptimization Model Example Optimize bandwidth partition among bands, given pricing, costs

The optimal band partitioning problem (OBP) can be defined as finding and values that satisfy:

Decision Variables: : partitioned capacity for band b in link j

: 1 if node pair p, traffic pair b, goes through route r , and 0 otherwise

Other Parameters: : effective bandwidth derived based on on-line measurement and QoS.

Refer to the paper for other parameters.

bjC pbrx

Db Jj

bjbj wCmin

Pp Rr

bjpbrpbrjpbpb

pb

CxAEv JjDb ,

Db

jbj CAPC Jj

pbRr

pbrx 1

pbbjpbr RraboveasjbpCorx ,,,,0,10

PpDb ,

bjC

pbrx

pbE

8

Results: Data PartResults: Data Part Systematic method comparison (pros and cons, complexity) Single node simulation: efficiency vs. QoS

SRD and LRD traffic (investigated several generators). Selected algorithms (e.g., Gaussian, Norros, Courcoubetis, DRDMW). Showed gains, detailed statistics, satisfied QoS while saving bandwidth. Proposed novel dynamic time scale technique.

Established Linux QoS test-bed, with MPLS, LDP, etc. Ported algorithms to IETF RTFM Confirmed simulation results with realistic emulation

9

Measurement Methods: CompareMeasurement Methods: Compare

10

Emulation: Emulation: RTFMRTFM

EB estimation was also implemented in the meter side. This eliminated the need for large SNMP transfers and resulted in faster response.

11

Emulation: Bandwidth SavingsEmulation: Bandwidth Savings

12

Emulation: Preservation of QoSEmulation: Preservation of QoS