Transcript Next Generation Internet

NEXT GENERATION INTERNET
PHYSICAL TESTBED:
RESEARCH AND MANAGEMENT ISSUES
Ian F. Akyildiz
C. Scoglio, J. de Oliveira, T. Anjali, L. Chen,
J. A. Smith*, G. Uhl* and A. Sciuto*
Broadband and Wireless Networking Laboratory
School of Electrical and Computer Engineering
Georgia Institute of Technology, Atlanta, GA, USA
*NASA Goddard Space Flight Center
Greenbelt, MD, 20771, USA
IP QoS NETWORK PROJECT
(NASA Goddard, Raytheon, Swayles)
 Challenges
– Differentiated
Services
NASA Ames
Research
– End-to-End
Center
QoS
– Integrated
Services for
Multimedia
– Network
Management
IFA'02
Abilene
NASA
Goddard Space
Flight Center
BWN
Laboratory
GATECH
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
2
BWN-Lab Physical Testbed:
Experiments and Issues
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
3
BWN-Lab TESTBED
Lightstream
1010
ATM
622 Mbps
7505
Gigabit Ethernet
Fast Ethernet
7204 VXR
ATM
155 Mbps
Catalyst
6506
Catalyst
4000
7204 VXR
External
Lightstream
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
4
TESTBED HARDWARE
 2 Cisco 7200 routers:
– FastEthernet/OC3/GigabitEthernet interfaces
 Cisco 7500 router:
– GEIP+/OC12 interfaces
 Cisco Catalyst 6506 layer 3 switch:
– GEIP+ interface
 Cisco Catalyst 4000 switch:
– FastEthernet ports
 Cisco LightStream 1010 switch:
– OC3 interfaces
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
5
TESTBED SOFTWARE
 Cisco IOS
– Version 12.2(1)E1 on the routers
– Native IOS on switch 6500
 End-hosts
– MS Win ME and Linux RedHat 7.2
– ALTQ 2.2: Scheduling/Queueing Software
– Iperf: traffic generation and traffic
characteristic measurement
– MRTG and MRTG++ (modified MRTG for 10s
sampling)
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
6
DiffServ TESTBED
NETWORK TOPOLOGY
DS domain
NASA
Abilene
marking
scheduling policing
DS domain
shaping
BWN-Lab
marking
DS domain
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
7
DiffServ EXPERIMENTS
 EDGE COMPONENTS
– Classification/Marking: Policy Based Routing (PBR)
allows classifications based on IP Precedence.
– Policing: Committed Access Rate (CAR) enforces a
specified traffic profile preventing non-conformant
traffic from entering the network.
– Shaping: Generic Traffic Shaping (GTS) follows the
token bucket algorithm.
 CORE COMPONENTS
– Queueing: Class-Based WFQ (CBWFQ) regulates
traffic submitted to the network, which may delay
packets to adjust traffic stream characteristics to a
defined profile.
– Congestion avoidance: Weighted RED (WRED) allows
definition of multiple drop probability profiles.
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
8
DiffServ EXPERIMENTS
SCHEDULING WITH CBWFQ
Class
Time of Start
Requested BW (Mbps)
Minimum Guaranteed BW (Mbps)
EF (UDP)
0
50
30
AF4 (UDP)
9
40
25
AF1 (UDP)
5
40
20
14
40
None
BE (UDP)
60
Bandwidth (Mbps)
50
40
EF
AF4
30
AF1
20
BE
10
0
1
3
5
7
9
11
13
15
17
19
Time (sec)
We validated CAR and CBWFQ as the policing and scheduling
mechanisms for DiffServ implementation
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
9
MPLS TESTBED
NETWORK TOPOLOGY
Abilene
rtr3
7505
Gigabit
rtr1
FastEthernet
7200
MPLS Tunnel 1
MPLS Tunnel 2
MPLS Tunnel 3
Gigabit
LAN 1
IFA'02
rtr2
7200
Gigabit
LAN 2
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
10
MPLS TE EXPERIMENTS
 Goal: Evaluate the benefits of MPLS TE
– Case Study 1: Traditional IP Network (Min Hop  rtr1-rtr2)
– Best-effort service only
– Two 40 Mbps UDP flows are sent from rtr1 to rtr2
– Two 100 Mbps TCP flows are sent from rtr1 to rtr2
– All flows take the min-hop path (FastEthernet) and are limited
to a total of 100 Mbps. UDP starves the TCP flows.
50
Bandwidth (Mbps)
45
40
35
TCP1
30
TCP2
25
UDP1
20
UDP2
15
10
5
19
17
15
13
11
9
7
5
3
1
0
Time (sec)
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
11
MPLS TE EXPERIMENTS
 Goal: Evaluate the MPLS TE Properties
– Case Study 2: MPLS Network – Mixed Flows
– 3 MPLS tunnels were set up.
– Two 40 Mbps UDP flows sent from rtr1 to rtr2
– Two 100 Mbps TCP flows sent from rtr1 to rtr2
– Tunnel1: UDP1 + TCP1; Tunnel2: TCP2; Tunnel3: UDP2
– TCP1 reduces rate when UDP1 arrives due to BW
contention
100
90
80
70
UDP1
60
TCP1
50
TCP2
40
UDP2
30
20
10
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
19
17
15
13
11
9
7
5
3
1
0
12
MPLS TE EXPERIMENTS
 Goal: Evaluate the MPLS TE properties
– Case Study 3: MPLS Network – Separate Flows
– 3 MPLS tunnels
– Two 40Mbps UDP flows sent from rtr1 to rtr2
– Two 100 Mbps TCP flows sent from rtr1 to rtr2
– Tunnel1: TCP1; Tunnel2: TCP2; Tunnel3: UDP1 + UDP2
– No interference between TCP and UDP
100
90
80
70
TCP1
60
TCP2
50
UDP1
40
UDP2
30
20
10
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
19
17
15
13
11
9
7
5
3
1
0
13
EXPERIMENTAL
CONCLUSIONS
 The MPLS TE provides better resource
utilization and throughput
– Cisco’s MPLS tunnels implementation does
not enforce the limit on the tunnel reserved
bandwidth – needs improvement
– CAR policing is not implementable on
Tunnel interfaces
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
14
DiffServ-AWARE TE
PREEMPTION EXPERIMENTS
 Goal: To evaluate Cisco’s preemption policy
– 3 MPLS tunnels were setup between routers, sharing a
FastEthernet link
– Tunnels 1, 2, and 3 together require the total link
bandwidth
– A new bandwidth request arrives for Tunnel 4, which
has higher priority than the other 3 tunnels
– One of the previously established tunnels must be
preempted. Which one?
New Tunnel
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
15
DiffServ-AWARE TE
PREEMPTION EXPERIMENTS
 Cisco’s preemption policy:
– Tunnel priority: lowest priority (numerically higher)
– Tunnel age: Tunnel created earliest
Lowest Priority
Tunnel
Oldest Tunnel
Priority Bandwidth Preempted
Priority Bandwidth Preempted
Tunnel 1
7
50
50
Tunnel 2
7
30
6
20
Tunnel 3
6
20
1
20
Tunnel 4
1
20
Tunnel 1
7
30
Tunnel 2
7
Tunnel 3
Tunnel 4
IFA'02
Tunnel
X
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
X
16
EXPERIMENTAL
CONCLUSIONS
 DiffServ-aware TE support in Cisco’s IOS is not
completely deployed
– Preemption is purely based on tunnel
priority and age – waste of resources
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
17
TEAM/AA Architecture:
Managing Multiple Domain
DiffServ MPLS Networks
Research Contributions and Issues
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
18
RELATED WORK
 RATES (Routing and Traffic Engineering Server)
– Software by Bell Labs for MPLS Traffic Engineering (TE)
– Uses Common Open Policy Service (COPS) and Minimum
Interference Routing Algorithm (MIRA)
– It achieves TE by routing of bandwidth guaranteed LSPs
 TEQUILA (TE for QoS in the Internet at Large Scale)
– European research project for end-to-end QoS in DiffServ
network
– Components for monitoring, TE, SLS management, and
policy management
– Algorithms and techniques are not concretely defined yet
and their quantitative evaluation has not been carried out
 MATE (Multipath Adaptive Traffic Engineering)
– Software by Bell Labs for MPLS TE
– Assumes LSP layout using a long term traffic matrix. The
focus is on load balancing short term traffic fluctuations
– Not designed for bandwidth guaranteed services
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
19
TRAFFIC ENGINEERING AUTOMATED
MANAGER (TEAM)
“Design and Management Tools for an MPLS Domain QoS Manager,”
to appear in Proceedings of SPIE ITCOM 2002, Boston, August 2002.
Intra-domain operation
To Neighboring
TEAM
To Neighboring
Domain
Management
Plane
LSP Routing
Traffic Routing
LSP Setup/
Dimensioning
DiffServ/MPLS
Domain
LSP Capacity
Allocation
LSP Preemption
Location
Management
Handoff
Management
TEAM
IFA'02
Network
Planning
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
20
TEAM/AA
ARCHITECTURE
Inter-domain operation
TEAM
TEAM
TEAM
AA
AA
AA
AA
AA
AA
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
21
TEAM/AA
ARCHITECTURE
 Traffic Engineering Automated
Manager (TEAM) and Adaptive Agent
(AA):
– Manage heterogeneous networks
– Different services such as best-effort, real-time,
etc.
– Different network technologies such as wired and
wireless mobile networks
– Manage large networks
– Multiple domains
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
22
TEAM COMPONENTS
 Traffic Engineering Tool [5]
– Resource Management
–Optimal Policy for LSP Setup [1, 2]
– Adaptive preemption policy for LSPs [3]
– Traffic estimation and resource allocation scheme
[4,6]
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
23
RESOURCE MANAGEMENT – LSP SETUP
“A New Threshold-Based Policy for Label Switched Path Setup in MPLS Networks,”
in proceedings of ITC 2001, Salvador da Bahia, Brazil, pp. 1-11, December 2001.
“Optimal Policy for Label Switched Path Setup in MPLS Networks,”
accepted for publication in Computer Networks Journal, 2002.
Determine an Adaptive Traffic Driven Policy for LSP
Setup and Dimensioning for each MPLS Network.
 Based on Markov Decision Process theory.
 Objective Function:
– Minimize the expected infinite-horizon
discounted total cost.
 To determine the optimal policy,  the
transition probabilities and the optimality
equations
 The optimality equations are solved using the
Value Iteration Algorithm.
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
24
OPTIMIZATION PROBLEM
Optimal policy * such that
Optimality equation
where
IFA'02
wb ( S , a )  w sw ( S , a )
r ( S , a )  Wsign( S , a ) 
  
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
25
TEAM COMPONENTS
 Traffic Engineering Tool [5]
– Resource Management
– Optimal Policy for LSP Setup [1, 2]
–Adaptive Preemption Policy for LSPs [3]
– Traffic estimation and resource allocation scheme [4,6]
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
26
RESOURCE MANAGEMENT –
LSP PREEMPTION
“A New Preemption Policy for DiffServ-Aware Traffic Engineering to Minimize Rerouting,” to
appear in Proceedings of IEEE INFOCOM 2002, New York City, June 2002.
 Non-real time applications may afford
to have their transmission rate reduced.
 By reducing the rate in a fair fashion:
– These LSPs would not be torn down,
– There would be no service disruption,
extra setup and tear down signaling
– THERE WOULD BE NO REROUTING
DECISIONS
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
27
ADAPTIVE PREEMPTION POLICY
(Contd.)
 Combines the three main
preemption criteria:
– Priority of preempted LSPs
– Number of preempted LSPs
– Bandwidth of preempted LSPs
 Optimization formulation and
heuristic
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
28
ADAPTIVE PREEMPTION POLICY –
OPTIMIZATION FORMULATION
Minimize:
F =  (priority cost) +  (number of LSPs)
+  (preempted bandwidth) + BW module cost
Subject to:
– Number of preempted modules  r
– Number of preempted modules in a preempted LSP is
equal to total number of modules in the LSP.
– Number of preempted modules in a rate reduced LSP is
less than  % of the total number of modules in the LSP.
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
29
PERFORMANCE COMPARISON:
COMMERCIAL VERSUS ADAPTIVE POLICY
Number of Preempted LSPs
IFA'02
Bandwidth Wastage
Requested bw
Mbps
Commercial
Policy
Adaptive
Policy
Requested bw
Mbps
Commercial
Policy
Adaptive
Policy
15
1
0
15
10
0
25
1
0
25
0
0
30
2
0
30
70
0
40
2
0
40
60
0
55
2
0
55
45
0
65
2
0
65
35
0
90
2
1
90
10
0
100
2
2
100
0
0
155
6
2
155
16
0
185
7
3
185
71
0
240
7
4
240
16
0
280
8
5
280
1
0
325
10
6
325
66
0
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
30
TEAM COMPONENTS
 Traffic Engineering Tool [5]
– Resource Management
– Optimal Policy for LSP Setup [1, 2]
– Adaptive Preemption Policy for LSPs [3]
–Traffic Estimation and Resource
Allocation Scheme [4,6]
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
31
RESOURCE MANAGEMENT –
LSP BANDWIDTH ALLOCATION
A method to determine Bandwidth Allocation for
LSPs with less bandwidth wastage and less redimensioning in an MPLS Network.
 Simple method is over-provisioning or cushion
 New Method  based on Kalman filter for
optimal estimation of the traffic and capacity
prediction by determining transition
probabilities
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
32
TRAFFIC ENGINEERING AUTOMATED
MANAGER (TEAM)
“Design and Management Tools for an MPLS Domain QoS Manager,”
to appear in Proceedings of SPIE ITCOM 2002, Boston, August 2002.
Intra-domain operation
To Neighboring
TEAM
To Neighboring
Domain
Management
Plane
LSP Routing
Traffic Routing
LSP Setup/
Dimensioning
DiffServ/MPLS
Domain
LSP Capacity
Allocation
LSP Preemption
Location
Management
Handoff
Management
TEAM
IFA'02
Network
Planning
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
33
TEAM/AA
ARCHITECTURE
Inter-domain operation
TEAM
TEAM
TEAM
AA
AA
AA
AA
AA
AA
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
34
PUBLICATIONS
[1] C. Scoglio, T. Anjali, J. de Oliveira, I. Akyildiz, and G. Uhl,
“A New Threshold-Based Policy for Label Switched Path Setup in
MPLS Networks,” in proceedings of ITC 2001, Salvador da Bahia,
Brazil, pp. 1-11, December 2001.
[2] T. Anjali, C. Scoglio, J. de Oliveira, I. Akyildiz, and G. Uhl, “Optimal
Policy for Label Switched Path Setup in MPLS Networks,” accepted
for publication in Computer Networks Journal, 2002.
[3] J. de Oliveira, C. Scoglio, I. Akyildiz, and G. Uhl, “A New Preemption
Policy for DiffServ-Aware Traffic Engineering to Minimize
Rerouting,” to appear in proceedings of IEEE INFOCOM 2002, New
York City, June 2002.
[4] C. Bruni, C. Scoglio, and S. Vergari, “Optimal Capacity Provisioning
for Label Switched Paths in MPLS Networks,” to appear in
proceedings of IFIP-TC6 Networking 2002, Pisa, Italy, May 2002.
[5] J. de Oliveira, C. Scoglio, T. Anjali, L. Chen, I. Akyildiz, and G. Uhl,
“Design and Management Tools for an MPLS Domain QoS Manager,”
to appear in Proceedings of SPIE ITCOM 2002, Boston, August 2002.
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
35
PUBLICATIONS (Contd.)
[6] T. Anjali, C. Scoglio, I. Akyildiz, and G. Uhl, “A New Scheme
for Traffic Estimation and Resource Allocation for Bandwidth
Brokers,” submitted for publication, 2002.
[7] T. Anjali, C. Scoglio, L. Chen, I. Akyildiz, and G. Uhl, “ABEst:
An Available Bandwidth Estimator within an Autonomous
System,” submitted for publication, 2002.
[8] J. de Oliveira, F. Martinelli, and C. Scoglio, “SPeCRA: A
Stochastic Performance Comparison Routing Algorithm for
LSP Setup in MPLS Networks,” submitted for publication,
2002.
[9] J. L. Marzo, E. Calle, C. Scoglio, and T. Anjali, “Adding QoS
Protection in Order to Enhance MPLS QoS Routing,”
submitted for publication, 2002.
IFA'02
NGI PHYSICAL TESTBED: RESEARCH AND MANAGEMENT ISSUES
36