Upload
jeffry-hudson
View
214
Download
2
Embed Size (px)
Citation preview
DiffServ/MPLS Network Design and
Management
Doctoral Dissertation
Tricha AnjaliBroadband and Wireless Networking Laboratory
Advisor: Dr. Ian F. Akyildiz
March 30, 2004 2BWN Lab - Tricha Anjali
Contents
• Introduction• Network Management• TEAM Structure• LSP/SP Setup • Traffic Routing• Available Bandwidth Estimation• End-to-end Available Bandwidth Measurement• Inter-domain Management• TEAM Implementation• Conclusions• Future Work
March 30, 2004 3BWN Lab - Tricha Anjali
Goals
• Two-fold which are complementary:
– Guarantee Quality of Service for the required applications.
– Use the network resources efficiently.
March 30, 2004 4BWN Lab - Tricha Anjali
MultiProtocol Label Switching
• Explicitly routed point-to-point paths called Label Switched Paths (LSPs)
• Support for traffic engineering and fast reroute
• Simpler switching operations
March 30, 2004 5BWN Lab - Tricha Anjali
Generalized MPLS
• GMPLS is a set of protocols for a common control of packet and wavelength domains
• Reserve a wavelength on a path (Lambda Switched Path or SP) for an aggregation of flows
srcdest
March 30, 2004 6BWN Lab - Tricha Anjali
DiffServ + GMPLS
• DiffServ – Scalable service differentiation
• DiffServ + GMPLS – Class differentiation for QoS
provisioning– Traffic Engineering for DiffServ classes
for efficient use of resources
March 30, 2004 7BWN Lab - Tricha Anjali
Network Model
Class Type 0 (BE)
Class Type 1 (AF)
Class Type 2 (EF)
MPLS Networks Link: Label Switched Path (LSP)
Optical NetworkLink: fiber
Wavelength NetworkLink: lambda Switched Path (SP)
March 30, 2004 8BWN Lab - Tricha Anjali
MPLS Network Management
• Existing MPLS network management tools:
– RATES (Bell Labs, 2000):✓ Sets up bandwidth guaranteed LSPs✘ Does not support DiffServ✘ No performance measurement and analysis
– DISCMAN (EURESCOM, 2000): Provides test and analysis results of DiffServ and MPLS-
based DiffServ✘ Does not provide its own management system
functionality
March 30, 2004 9BWN Lab - Tricha Anjali
MPLS Network Management
• Other existing MPLS network management tools:
– MATE (Bell Labs, Univ. Michigan, Caltech, Fujitsu, 2001): The goal is to distribute the traffic across several LSPs
established between a given ingress and egress node pair✘ Not for traffic that requires bandwidth reservation
– TEQUILA (European Union Project, 2002): Global and integrated approach to network design and
management✘ No network management methods developed and
implemented✘ No evaluation of performances
March 30, 2004 10BWN Lab - Tricha Anjali
A New Network Management Tool
• Traffic Engineering Automated Manager (TEAM) – Automated– Monitors the network performance – Implements various algorithms for
handling events in MPLS and optical network
– Allows efficient use of resources and prompt responses
March 30, 2004 11BWN Lab - Tricha Anjali
Big Picture of TEAMTraffic Engineering Automated Manager
Rou
teR
esou
rce
LSP Routing
Traffic Routing
LSP Preemption
LSP/SP Setup/Dimensioning
Management Plane DiffServ/
GMPLS Domain
SimulationTool (ST)
TrafficEngineeringTool (TET)
Measurement/Performance
EvaluationTool (MPET)
TEAM
To neighboring TEAM
Network Dimensioning and Topology Design
March 30, 2004 12BWN Lab - Tricha Anjali
LSP and SP Setup Problem
Find an adaptive traffic driven policy for dynamic setup and tear-down of LSPs and SPs.
Why not the fully connected topology?Too many LSPs for increasing number of routers N (N2 problem)
Why not a fixed topology?Because traffic is unpredictable
- “Optimal Policy for LSP Setup in MPLS Networks,” Computer Networks Journal, June 2002- “LSP and SP Setup in GMPLS Networks,” Proceedings of IEEE INFOCOM, March 2004
March 30, 2004 13BWN Lab - Tricha Anjali
LSP and SP Setup Problem• Arrival of bandwidth request
• Decision among: – Option 1: no action– Option 2: setup a direct LSP– Option 3: setup a direct SP and LSP
srcdest
1
2
3
March 30, 2004 14BWN Lab - Tricha Anjali
LSP and SP Setup
• Optical network virtual topology design algorithms– Chen 1995, Davis 2001, Krishnaswamy
2001: Design the network off-line with a given traffic matrix
– Gençata 2003 : On-line virtual topology adaptation approach for optical networks✘Does not combine optical and MPLS layers
March 30, 2004 15BWN Lab - Tricha Anjali
Assumptions• Routing Assumption
– Default topologies
– Packets are routed either on• the direct LSP(i,j) or • the min-hop path P(i,j) over the default MPLS network
– LSPs are routed either on• the direct SP or• the min-hop path P
ij over the default optical network
– a new LSP can not be routed on a previously established non-default SP
March 30, 2004 16BWN Lab - Tricha Anjali
Model Formulation
• Events and Decision Instants
– MPLS network• Arrival/Departure of bandwidth requests
between (i, j)
– Optical network• Arrival of LSP(i, j) capacity increment/decrement
requests
March 30, 2004 17BWN Lab - Tricha Anjali
Model Formulation• State vector (local)
– MPLS network s = (A, Bl, Bp)• Available capacity (A)• Bandwidth requests on direct LSP (Bl) or
on min-hop path (Bp)
– Optical network s = (A, Bl, Bp, k)• Available capacity (A)• Capacity requests on direct SP (Bl) or on
min-hop path (Bp)• Number of SPs between the node pair (k)
March 30, 2004 18BWN Lab - Tricha Anjali
Model Formulation (Contd.)Action Variables
MPLS network
Optical network
1 setup or re-dimension LSP
0 no action on LSPa
1 setup or re-dimension SP
0 no action on SPa
March 30, 2004 19BWN Lab - Tricha Anjali
Cost ModelIncremental cost
W = Wb + Wsw+ Wsign
– Wb(s,a) : Bandwidth cost – Wsw(s,a) : Switching cost – Wsign(s,a) : Signaling cost if LSP/SP is set-up or
re-dimensioned
• Wb and Wsw are linear with respect to the bandwidth request and time
• Wsign is incurred only if the decision is a = 1
March 30, 2004 20BWN Lab - Tricha Anjali
Optimal Setup Policy
• Based on Markov Decision Process Theory
• Minimize expected infinite-horizon discounted total cost
• Determine transition probabilities and optimality equations
• Solve the optimality equations with value iteration algorithm
Optimal policy stationary control-limit
March 30, 2004 21BWN Lab - Tricha Anjali
Optimization (MPLS network)
( , ) ( , )( , ) ( , ) b sw
sign
w S a w S ar S a W S a
1
0
( )0
0
( ) ( , ) ( , ) ( , )m
m m
m
tt t t
S sign m b m sw mm t
v S E e W S a e w S a w S a dt
*
( ) inf ( )v s v s
_
( ) min ( , ) ( | , ) ( )a A
j S
v S r S a q j S a v j
Optimal policy * such that
Optimality equations
where
March 30, 2004 22BWN Lab - Tricha Anjali
Optimal Policy (MPLS Network)
* * * *{ , , , }d d d
*
**
*
, , ,0 for0
, , ,0 , , ,0 for, , ,1 , , ,1, , ,1 , , , 2
0 , , ,3
L P
L P L P
L P L P
L P L P
L P
S A B B A b
a A B B S A B B A ba A B Bd S A B Ba A B B S A B B
S A B B
* ** 1 , , 2 ,3 0, , ,3
, , ,00 otherwise
s a L P L PL P
c h c v A B B b v B B b ba A B B
* ** 1 , , ,3 0, , ,3
, , ,10 otherwise
s a L P L PL P
c h c v A b B b B b v B B b ba A B B
where
March 30, 2004 23BWN Lab - Tricha Anjali
Optimization (Optical Network)
( , ) ( , )( , ) ( , ) b sw
sign
w S a w S ar S a W S a
1
0
( )0
0
( ) ( , ) ( , ) ( , )m
m m
m
tt t t
S sign m b m sw mm t
v S E e W S a e w S a w S a dt
*
( ) inf ( )v s v s
_
( ) min ( , ) ( | , ) ( )a A
j S
v S r S a q j S a v j
Optimal policy * such that
Optimality equations
where
March 30, 2004 24BWN Lab - Tricha Anjali
Optimal Policy (Optical Network)
* * * *{ , , , }d d d
*
*
0,0, ,0,0
0 0,0, ,0,1
0 , , , ,0 0,
1 , , , ,0 0,
0 , , , ,1 0,
1 , , , ,1 0,
F
F
F F
F F
F F
F F
a S B
S B
S A B B k k A b Bd
S A B B k k A b B
S A B B k k A W b B
S A B B k k A W b B
* ** ' '
1 0,0, 2 ,0,1 2 , 2 ,0,1,1
0 otherwise
FcapF F F F
x y
c Whc c h v B b v W B b B b
a
where
March 30, 2004 25BWN Lab - Tricha Anjali
Sub-optimal Policy
• Optimal policy is difficult to pre-calculate because of large number of possible system states
• Sub-optimal policy that is fast and easy to calculate
• Minimizes the cost incurred between two decision instants
• Maintains the threshold structure of the optimal policy
March 30, 2004 26BWN Lab - Tricha Anjali
Sub-optimal Policy (MPLS)# # # #{ , , , }d d d
1
1#
1
, , ,0 for0
, , ,0 , , ,0 for, , ,1 , , ,1, , ,1 , , , 2
0 , , ,3
L P
L P L P
L P L P
L P L P
L P
S A B B A b
a A B B S A B B A ba A B Bd S A B Ba A B B S A B B
S A B B
1 1, , ,0
0 otherwiseP Th
L P
B b Ba A B B
1 1
, , ,10 otherwise
P ThL P
B Ba A B B
1s a
Th
ip mpls
c h cB
h c c
where
where
March 30, 2004 27BWN Lab - Tricha Anjali
Sub-optimal Policy (Optical)
# # # #{ , , , }d d d
1
*
0,0, ,0,0
0 0,0, ,0,1
0 , , , ,0 0,
1 , , , ,0 0,
0 , , , ,1 0,
1 , , , ,1 0,
F
F
F F
F F
F F
F F
a S B
S B
S A B B k k A b Bd
S A B B k k A b B
S A B B k k A W b B
S A B B k k A W b B
' '
1
( )( )1
( 1)( )
0 otherwise
F Fx y capF
Fopt
c c h c WhB b
a h c c
where
March 30, 2004 28BWN Lab - Tricha Anjali
Performance Evaluation
Example network:
• Network has 10 nodes and 17 links
• Cph = 1000 Mbps
• Diameter = length of longest shortest path = 3
March 30, 2004 29BWN Lab - Tricha Anjali
Comparison
Discount factor=0.5 Discount factor=0.1
Discounted total cost vs. Initial state
0
50
100
150
200
250
300
350
400
450
Initial State
Exp
ect
ed
To
tal C
ost
OptimalSub-optimal
[1,0,0] [1,1,1] [2,5,1] [1,5,1] [1,5,5] [1,5,7] [1,5,10] [1,10,7] [7,3,0]
[0,1,2] [3,8,0]
[1,1,7] [2,6,0]
[2,6,0] [1,5,1]
[1,5,1]
[4,2,0]
[4,2,0] [3,3,1]
[3,3,1] [8,4,0]
[2,4,0] [9,2,0]
[9,2,0]
0
200
400
600
800
1000
1200
1400
1600
Initial State
Exp
ect
ed
To
tal C
ost
OptimalSub-optimal
[1,0,0] [1,1,1] [3,5,1] [1,5,1] [1,5,5] [1,5,10] [1,5,7] [1,10,7]
[7,13,0]
[4,13,0] [10,10,0]
[9,11,0] [7,13,1]
[2,13,3] [8,12,0]
[9,8,0]
[4,16,0]
[8,10,0] [0,7,4]
[1,14,0] [1,5,11]
[1,5,7]
[1,5,3] [1,10,6]
March 30, 2004 30BWN Lab - Tricha Anjali
Experimental Results
What happens when we homogeneously increase traffic on selected node pairs
– LSPs with larger number of default LSPs in their path are established first
– SPs with larger number of default SPs that need re-dimensioning in their path are established first
March 30, 2004 31BWN Lab - Tricha Anjali
Heuristics for Comparison
Heuristic 1: Fully connected LSP network
Heuristic 2: LSP re-dimensioned exactly
Heuristic 3: LSP re-dimensioned with extra capacity
In each heuristic, SP network is fully connected
March 30, 2004 32BWN Lab - Tricha Anjali
Total Expected Cost
2 4 6 8 100
200
400
600
800
1000
1200
Experiment number
Exp
ect
ed
To
tal c
ost
OptimalSub-optimalHeuristic 1Heuristic 2Heuristic 3
March 30, 2004 33BWN Lab - Tricha Anjali
Bandwidth Wastage in MPLS Network
2 4 6 8 100
200
400
600
800
1000
Experiment number
Ma
x b
an
dw
idth
wa
sta
ge
in L
SP
s (M
bp
s)
Sub-optimalHeuristic 1Heuristic 2Heuristic 3
March 30, 2004 34BWN Lab - Tricha Anjali
Big Picture of TEAMTraffic Engineering Automated Manager
Rou
teR
esou
rce
LSP Routing
Traffic Routing
LSP Preemption
LSP/SP Setup/Dimensioning
Management Plane DiffServ/
GMPLS Domain
SimulationTool (ST)
TrafficEngineeringTool (TET)
Measurement/Performance
EvaluationTool (MPET)
TEAM
To neighboring TEAM
Network Dimensioning and Topology Design
March 30, 2004 35BWN Lab - Tricha Anjali
QoS Routing- “A New Path Selection Algorithm for MPLS Networks Based on Available Bandwidth
Estimation,” Proceedings of QoFIS, October 2002
- “Traffic Routing in MPLS Networks Based on QoS Estimation and Forecast,” submitted
Find a low cost feasible path for routing traffic flows in MPLS networks adaptively.
Why adaptive?Because MPLS network topology is changing
Existing routing algorithms• Heuristic solutions of the delay constrained least cost problem• LSP routing algorithms (MIRA, PBR)
March 30, 2004 36BWN Lab - Tricha Anjali
Routing Algorithm
• Notations– puv: path in the MPLS network
– puv= (lux, …, lzv)
– Alij/dl
ij: Available capacity/delay on lij
– npuv: Number of LSPs in puv
–
–
minij uv
p luv ij
l pA A
ij uv
p luv ij
l p
d d
March 30, 2004 37BWN Lab - Tricha Anjali
Cost ModelLSP cost
W = Wb + Wsw+ Wsign+WAB+Wd
– Wb and Wsw linear with respect to the bandwidth request and duration of request
– Wsign is instantaneous– WAB is inversely related to LSP available bandwidth– Wd linear with respect to delay on the LSP
Path cost
Wp = ∑ LSP costs + (n-1) ( Relay node cost )
March 30, 2004 38BWN Lab - Tricha Anjali
Routing Problem
Find the path such that
subject to feasibility constraints
** : minuv
p puv uv uv
pp W W
*
*
*
max
min
,
,
.
puv
puv
puv
n k
d d
A A
March 30, 2004 39BWN Lab - Tricha Anjali
Routing Algorithm
• Heuristic of the exact problem
• Path set size restricted to F
• Set populated by paths with increasing length
• Feasibility check
• Cost comparison
March 30, 2004 40BWN Lab - Tricha Anjali
Partial Information
• Estimation algorithm for accurate state information
• Linear prediction
• Dynamically change the number of past samples based on prediction performance
March 30, 2004 41BWN Lab - Tricha Anjali
Performance Evaluation
Popular ISP topology with link capacity = 155 c.u.
March 30, 2004 42BWN Lab - Tricha Anjali
Rejection Ratio
0 5 10 15 20 250
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Experiment
Per
cent
age
SPProposed
March 30, 2004 43BWN Lab - Tricha Anjali
Minimum Available Bandwidth
0 5 10 15 20 250
10
20
30
40
50
60
70
Experiment
Cap
acity
uni
ts
SPProposed
March 30, 2004 44BWN Lab - Tricha Anjali
Paths with Relay Nodes
0 5 10 15 20 250
20
40
60
80
Experiment
Num
ber
SPProposed
March 30, 2004 45BWN Lab - Tricha Anjali
Big Picture of TEAMTraffic Engineering Automated Manager
Rou
teR
esou
rce
LSP Routing
Traffic Routing
LSP Preemption
LSP/SP Setup/Dimensioning
Management Plane DiffServ/
GMPLS Domain
SimulationTool (ST)
TrafficEngineeringTool (TET)
Measurement/Performance
EvaluationTool (MPET)
TEAM
To neighboring TEAM
Network Dimensioning and Topology Design
March 30, 2004 46BWN Lab - Tricha Anjali
Available Bandwidth Measurement
Measure/estimate the available bandwidth in a link/path to analyze the performance of the network
Various existing tools to measure narrow link capacity– Pathchar based (Jacobson 1997) : link-by-link measurement– Packet pair based (Keshav 1991): end-to-end capacity– Nettimer (Lai 2001) : end-to-end capacity– AMP (NLANR 2002) : active link-by-link measurement– OCXmon (NLANR 2002): passive link-by-link measurement– MRTG (Oetiker 2000) : 5 min averages of link utilization– Pathload (Jain 2002): end-to-end available bandwidth
measurement
- “ABEst: An Available Bandwidth Estimator within an Autonomous System,” Proceedings of IEEE Globecom, November 2002
- “MABE: A New Method for Available Bandwidth Estimation in an MPLS Network,” Proceedings of IEEE NETWORKS, August 2002
March 30, 2004 47BWN Lab - Tricha Anjali
Available Bandwidth Estimator
• Assumptions– SNMP is enabled in the domain– MRTG++ is used to poll the network devices with
10 sec granularity• Notations
– L(t) : Traffic load at time t : Length of averaging interval of MRTG++– L[k] : Average load in [(k-1), k]– p : Number of past measurements in prediction– h : Number of future samples reliably predicted– Ah[k] : Available bandwidth estimate for [(k+1),
(k+h)]
March 30, 2004 48BWN Lab - Tricha Anjali
ABEst (Contd.)
• We use the past p samples to predict the utilization for the next h samples
• Utilize the covariance method for prediction
• Values of p and h varied according to the estimation error
kk-p+1 k+h
March 30, 2004 49BWN Lab - Tricha Anjali
ABEst (Contd.)
1. At time instant k, available bandwidth measurement is desired.
2. Find the vectors wa, a[1,h] using covariance method given p and the previous measurements.
3. Find and
4. Predict Ah[k] for [(k+1), (k+h)t].
5. At time (k+h)t, get
6. Find the error vector
7. Set k = k+h.
8. Obtain new values for p and h.
9. Go to step 1.
ˆ ˆ[ 1], , [ ] TL k L k h [ 1], , [ ] TL k p L k
[ 1], , [ ] Te k e k h
[ 1], , [ ] TL k L k h
March 30, 2004 50BWN Lab - Tricha Anjali
ABEst (Contd.)
• Covariance estimated as
• Covariance normal equations
• Ah[k] estimated– Either C – max{predicted utilization vector}– Or C – Effective bandwidth from the utilization vector
( , ) [ ] [ ]k
Li k N p
r n m L i n L i m
(0) (0, )(0,0) (0, 1)
(1) (1, )
( 1,0) ( 1, 1)( 1) ( 1, )
a LL L
a L
L La L
w r ar r p
w r a
r p r p pw p r p a
March 30, 2004 51BWN Lab - Tricha Anjali
ABEst (Contd.)
• Algorithm for h and p– If / > Th1, decrease h until hmin and increase p
till pmax multiplicatively
– If Th1 > /> Th2, decrease h until hmin and increase p till pmax additively
– If / < Th2, then:
• If > Th3*M2
E, decrease h until hmin and increase p till pmax
additively
• If Th3*M2E > > Th4*M2
E, keep h and p constant
• If < Th4*M2E, increase h and decrease p till pmin
additively
March 30, 2004 52BWN Lab - Tricha Anjali
Performance Evaluationhmin=10
200 300 400 500 60010
15
20
25
30
35
40
Sample number
Ba
nd
wid
th (
MB
/s)
Actual Peak-bw Est.Eff-bw Est.
March 30, 2004 53BWN Lab - Tricha Anjali
Performance Evaluation (Contd.)hmin=20
200 300 400 500 60010
15
20
25
30
35
40
Sample number
Ba
nd
wid
th (
MB
/s)
ActualPeak-bw Est.Eff-bw Est.
March 30, 2004 54BWN Lab - Tricha Anjali
End-to-end AB Measurement
• Motivation– Combine active and passive approaches– Most tools estimate narrow link capacity– Accuracy– Scalability– Statistical robustness– Not intrusive
- “TEMB: Tool for End-to-End Measurement of Available Bandwidth,” Proceedings of IEEE ELMAR, June 2003
March 30, 2004 55BWN Lab - Tricha Anjali
Tight Link Identification
• Measurement packets
• 10 measurement packets sent in a second, to make the tool non-intrusive
Version Type Length
Checksum
Data Record (optional)
Data Record (optional)
March 30, 2004 56BWN Lab - Tricha Anjali
Data Record
• Data record
• Inserted/modified by the hops of the path• Counter information from MIB-II in router
IP address
Counter
Timestamp
Speed
March 30, 2004 57BWN Lab - Tricha Anjali
Example of Auto-detection
SA.1.1.1 C.1.1.1
B.1.1.1
D.1.1.10 0checksum
8
D
0 0checksum
24
A.1.1.13245
23456310000000
0 0checksum
40
A.1.1.13245
23456310000000
C.1.1.12348754236
100000000 0checksum
8
0 0checksum
24
A.1.1.13272
23456810000000
0 0checksum
40
A.1.1.13272
23456810000000
C.1.1.12349854245
10000000
March 30, 2004 58BWN Lab - Tricha Anjali
Example of Non-min-hop Path
A.1.1.1 C.1.1.1
B.1.1.1
D.1.1.1
S D
0 1checksum
72
B.1.1.13245
23456310000000
D.1.1.12348754236
100000000C.1.1.1
532458643214
10000000
0 1checksum
72
B.1.1.13245
23456310000000
D.1.1.12348754236
100000000C.1.1.1
000
0 1checksum
72
B.1.1.13245
23456310000000
D.1.1.1000
C.1.1.1000
0 1checksum
72
B.1.1.1000
D.1.1.1000
C.1.1.1000
March 30, 2004 59BWN Lab - Tricha Anjali
Tight Link Identification
• 10 packets in one second• N packets back at source for analysis• Utilization of I-th interface at time tk
• Available bandwidth• At least agreelink of the estimates should concur
about the tight link identity.
Ik I IkA S U
( 1)
( 1)
for 2,3, ,Ik I k
Ikk k
k Nc c
Ut t
March 30, 2004 60BWN Lab - Tricha Anjali
Tight Link Identification (Contd.)
• All (N-1) estimates should be within [100, agreeavail]% of the minimum estimate
• Otherwise the next batch of 10 packets is sent.
• Average available bandwidth of interface I is
where n attempts have been made at measurement
( 1)
1
1
( 1)
n N
I Ikk
A An N
March 30, 2004 61BWN Lab - Tricha Anjali
MRTG-based Measurement
• More accurate estimation of tight link available bandwidth
• MRTG-based passive approach similar to ABEst
• Reliably predicts the utilization of the link for a future interval, that varies in size
March 30, 2004 62BWN Lab - Tricha Anjali
Big Picture of TEAMTraffic Engineering Automated Manager
Rou
teR
esou
rce
LSP Routing
Traffic Routing
LSP Preemption
LSP/SP Setup/Dimensioning
Management Plane DiffServ/
GMPLS Domain
SimulationTool (ST)
TrafficEngineeringTool (TET)
Measurement/Performance
EvaluationTool (MPET)
TEAM
To neighboring TEAM
Network Dimensioning and Topology Design
March 30, 2004 63BWN Lab - Tricha Anjali
Inter-domain Resource Management
• Inter-domain resource reservation agreements• Estimate the traffic on an inter-domain link and forecast its
capacity requirement, based on a measurement of the current usage
• Efficient resource utilization while keeping the number of reservation modifications to low values.
• Two approaches for resource allocation– Off-line : simple and predictable but lead to resource wastage– On-line : “Cushion” scheme (Terzis 2001) wherein extra
bandwidth is reserved over the current usage.• large number of re-negotiations to satisfy the QoS.
- “A New Scheme for Traffic Estimation and Resource Allocation for Bandwidth Brokers,” Computer Networks Journal, April 2003
- “Filtering and Forecasting Problems for Aggregate Traffic in Internet Links,” Performance Evaluation Journal, 2004
March 30, 2004 64BWN Lab - Tricha Anjali
Resource Reservation Problem
• Assumptions– Estimate traffic for one traffic class
– Number of established sessions is N and stays constant during analysis
– For each session, flows are defined as active periods
– Each flow has a constant rate of b bits per second
– Flows are assumed to be Poissonian with exponential inter-arrival times and durations
March 30, 2004 65BWN Lab - Tricha Anjali
Model Formulation
• Notations– y(m) : aggregate traffic on link at time m
– x(m) : number of active flows on link at time m
y(m) : noisy measure of the aggregate traffic on link at time m
– x(m) : estimate of x(m)
– pk(t) : probability that number of active flows at time t is k
March 30, 2004 66BWN Lab - Tricha Anjali
Traffic Estimation
• Generating function G(z,t), with the initial condition G(z,mT)=zx(m)
( )( )
( )
( 1)( , ) ( , ) where ( , )
( , ) ( 1)
N tx m
t
z z z eG z t C z t C z t
C z t z z e
ˆ ˆ ˆ( ) ( 1) ( ) ( ) ( 1)x m Ax m B k m y m CAx m CB
where ( )
( )1
( ) is Kalman Filter Gain
T
T
A e
NB e
C b
k m
March 30, 2004 67BWN Lab - Tricha Anjali
Allocation Forecasting• x(m) to forecast R(m+1)
• Define and Q as the transition probability matrix
0 1( ) ( ) ( )T
NP p t p t p t
1
1 1
( 1)
0 1
ˆ[ ( ), ]
and
where
1 1Define
Define ( ) min . .
Then ( 1) ( )
t mTmT
m TTt
N
mT
xx x m N
P QP Q Y Y
P Ye C C e Y P
P Y e dt C p p pT T
x m x s t p
R m bx m
March 30, 2004 68BWN Lab - Tricha Anjali
Performance EvaluationN=20, ==0.005
0 3000 6000 90000
5
10
15
20
25
Time (sec)
Ba
nd
wid
th (
Mb
ps)
ActualEPABBCushion
March 30, 2004 69BWN Lab - Tricha Anjali
Performance Evaluation (Contd.)
0 3000 6000 90000
5
10
15
20
25
Time (sec)
Ba
nd
wid
th (
Mb
ps)
ActualEPABBGaussian
March 30, 2004 70BWN Lab - Tricha Anjali
Big Picture of TEAMTraffic Engineering Automated Manager
Rou
teR
esou
rce
LSP Routing
Traffic Routing
LSP Preemption
LSP/SP Setup/Dimensioning
Management Plane DiffServ/
GMPLS Domain
SimulationTool (ST)
TrafficEngineeringTool (TET)
Measurement/Performance
EvaluationTool (MPET)
TEAM
To neighboring TEAM
Network Dimensioning and Topology Design
March 30, 2004 71BWN Lab - Tricha Anjali
TEAM Implementation
• TEAM has been implemented to run on a computer with the Linux OS.
• This testbed has been used as the platform to implement and test the operation of TEAM.
March 30, 2004 72BWN Lab - Tricha Anjali
TEAM Top-level Design
User interfaceserver
Commands
New bandwidthrequest
LSP Setup
Configurerouters
Routers
Trigger receiver
Configuration
Topology updates
Update topology
Preemption Reroute
Route
Create/Destroy/Resize LSP
Create/Resize LSP
Route
Label, path,
priority, bandwidth
Path, priority, bandwidth
LSPs to be destroyed
LSPs to be re-routed
Route
Path, priority,bandwidth
Topology change
New bandwidth
request
Path, priority,bandwidth
Label, path
Scheduler MRTGInterface Measurements
March 30, 2004 73BWN Lab - Tricha Anjali
TEAM Module Hierarchy
ABEST
COMMAND CONFIG
EVENTS
GRAPH
LSP_DB
LSP_SETUP
MPET
MRTG
PREEMPT
REQUEST_DB
REQUEST
REA
ROUTING
RRDTOOL
SCHEDULER
SNMP
TOPOLOGY
UI-SERVER
UI-PROTOCOL
RE-ROUTE
GSL
TET
NET_SNMP
March 30, 2004 74BWN Lab - Tricha Anjali
Performance Evaluation
• Topology with 40 nodes and 64 links of capacity 600 Mbps
• Comparison with a traditional manager– Shortest path routing for LSPs– Shortest path routing for traffic– LSP setup based on service level agreements– No LSP preemption– No on-line network measurements
March 30, 2004 75BWN Lab - Tricha Anjali
Generalized Medium Traffic Load
0 5 10 15 20 250
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Experiment
Rat
io
TMTEAM
Rejection Ratio
March 30, 2004 76BWN Lab - Tricha Anjali
Generalized Medium Traffic LoadMinimum AB Average AB
0 5 10 15 20 250
100
200
300
400
500
600
Experiment
Cap
acity
uni
ts
TMTEAM
0 5 10 15 20 25100
110
120
130
140
150
160
Experiment
Cap
acity
Uni
ts
TMTEAM
March 30, 2004 77BWN Lab - Tricha Anjali
Focused High Traffic LoadPriority 0 Rejection Priority 1 Rejection
0 5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
Experiment
Rat
io
TMTEAM
0 5 10 15 20 250
0.05
0.1
0.15
0.2
0.25
Experiment
Rat
io
TMTEAM
March 30, 2004 78BWN Lab - Tricha Anjali
Conclusions
Development of TEAM, an automated manager for MPLS networks, that performs network design and adaptive network management including LSP and traffic routing, LSP setup and capacity allocation, etc. based on network measurements.
March 30, 2004 79BWN Lab - Tricha Anjali
Future Work
• Heterogeneous large network management
• MPLS in Wireless Networks
• Network Tomography
March 30, 2004 80BWN Lab - Tricha Anjali
Publications1. “Building an IP Differentiated Services Testbed,”
Proceedings of IEEE ICT, June 2001
2. “A New Threshold-Based Policy for Label Switched Path Setup in MPLS Networks,” Proceedings of 17th ITC, September 2001
3. “Optimal Policy for LSP Setup in MPLS Networks,” Computer Networks Journal, June 2002
4. “Design and Management Tools for an MPLS Domain QoS Manager,” Proceedings of SPIE ITCOM, July 2002
5. “MABE: A New Method for Available Bandwidth Estimation in an MPLS Network,” Proceedings of IEEE NETWORKS, August 2002
6. “A New Path Selection Algorithm for MPLS Networks Based on Available Bandwidth Estimation,” Proceedings of QoFIS, October 2002
7. “ABEst: An Available Bandwidth Estimator within an Autonomous System,” Proceedings of IEEE GLOBECOM, November 2002
8. “A New Traffic Engineering Manager for DiffServ/MPLS Networks: Design and Implementation on an IP QoS Testbed,” Computer Communications Journal, March 2003
9. “A New Scheme for Traffic Estimation and Resource Allocation for Bandwidth Brokers,” Computer Networks Journal, April 2003
10. “Adding QoS Protection in Order to Enhance MPLS QoS Routing,”Proceedings of IEEE ICC, May 2003
March 30, 2004 81BWN Lab - Tricha Anjali
Publications (Contd.)11. “TEMB: Tool for End-to-End Measurement of Available Bandwidth,”
Proceedings of IEEE ELMAR, June 2003
12. “QoS On-line Routing and MPLS Multilevel Protection: A Survey,”IEEE Communications Magazine, October 2003
13. “Optimal Filtering in Traffic Estimation for Bandwidth Brokers,” Proceedings of IEEE GLOBECOM, December 2003
14. “LSP and SP Setup in GMPLS Networks,”Proceedings of IEEE INFOCOM, March 2004
15. “Threshold-Based Policy for LSP and SP Setup in GMPLS Networks,” Proceedings of IEEE ICC, June 2004
16. “New MPLS Network Management Techniques Based on Adaptive Learning,”IEEE Transactions on Neural Networks, 2004
17. “Filtering and Forecasting Problems for Aggregate Traffic in Internet Links,”Performance Evaluation Journal, 2004
18. “Traffic Routing in MPLS Networks Based on QoS Estimation and Forecast,” submitted for publication
19. “TEAM: A Traffic Engineering Automated Manager for DiffServ-based MPLS Networks,”submitted for publication