GMPLS Actively Managed WDM Testbed · GMPLS Actively Managed WDM Testbed Y. J. (Ray) Chen...
Transcript of GMPLS Actively Managed WDM Testbed · GMPLS Actively Managed WDM Testbed Y. J. (Ray) Chen...
UMBCLTS Review 2003 p1
GMPLS Actively Managed WDM Testbed
Y. J. (Ray) ChenDepartment of CSEE, UMBC
UMBCLTS Review 2003 p2
ObjectivelActively managed GMPLS Network
– Actively managed reconfigurable WDM optical network» Ultra-high capacity (~10Tps)» Scalability and efficiency of bandwidth utilization» Reconfiguration is buit into the traffic engineering process
– Delivery of quality of service (QoS)» Capable of providing variable levels of QoS according to service level
agreements (SLAs)– Considerable resiliency
» Prompt detection of failures» Fast protection switching/restoration
– Advanced network control and management
lBase on optical Ethernet – Internet traffic is dominated by Ethernet packets – Introduces QoS traffic engineering and Differentiated service– Low cost line rate and Ethernet switch devices– Very efficient packet delivery
UMBCLTS Review 2003 p3
Current GMPLS and Optical Ethernet Approachesl Inefficiency of current GMPLS implementation
– All-optical routing in the core region of the network» Coarse traffic granularity» High traffic blocking rate» Limited reconfiguration capability» No grooming; bandwidth utilization efficiency is limited
– MPLS and optical layers are managed separately» Complicates the OAM&P
– Layer 2 is passive and non-reconfigurable
l Difficulties of current Optical Ethernet– No carrier-class QoS– Lack of efficient OAM&P methods– Inefficient protection/restoration schemes
l Existing improvement attempts on Optical Ethernet– Resilient Packet Ring (RPR)
» Specially designed MAC for packet transport over fiber-optic rings» Introduces SONET-like protection and restoration schemes to optical Ethernet» Not suitable for other regular topologies, e.g., mesh topology
– Layer-3 switching» Merges layer 2 switching and layer 3 routing functions to a single switch box» Utilizes reconfigurable features of layer 2 to some extent
OXC
OXC OXC
OXC
OXC
OXC
O X C
Optical LayerOXC
MPLS Core Router
OXC/OADM
MPLS Edge Router
MPLS Layer
Combination of GMPLS and optical Ethernet–Incorporates both QoS and resilience features of GMPLS and efficiency of Ethernet
UMBCLTS Review 2003 p4
Our GMPLS Testbed
ü Implements an Integrated Reconfigurable Ethernet (layer 2) + Optical Switching Layer
− Even with legacy Ethernet switches (routing and signaling protocols required)
− Integrated management of the optical layer and Ethernet layer through QoS traffic engineering
üEnables traffic grooming and O/E/O wavelength conversion at the core of the network
üWavelength information and MAC addresses are utilized collectively to perform the GMPLS “label”functions
UMBCLTS Review 2003 p5
System Architecture
Ethernet Layer
OXC
OXC OXC
OXC
OXC
OXC
OXC
Optical Layer
OXC
MPLS Core Router
OXC/OADM
MPLS Edge Router
IP/MPLS Layer
Ethernet switch
â The label structureλ MAC Label EXP S TTL
L1 Label L2 Label MPLS Label
ã Three forwarding schemesall-opticalO/E/O at Ethernet layerO/E/O at IP/MPLS layer
– Different forwarding scheme can have different QoS level, e.g., bandwidth, latency, jitter, etc.
OXC
UMBCLTS Review 2003 p6
Benefitsl QoS path provisioning
– The device performance/response parameters are calibrated for each QoS level
– QoS paths are set up according to the committed service level agreement (SLA)
– Paths can be adjusted, i.e., reconfigured, according to traffic engineering decision
l QoS path performance improvement– Path latency reduction
l Efficient bandwidth utilization through O/E/O wavelength conversion and traffic grooming
l Potential system cost reduction– Inexpensive Ethernet switch devices– Low cost Ethernet deployment– Fast service provisioning
UMBCLTS Review 2003 p7
Hardware subjects for UMBC MPλS Testbed
UMBCLTS Review 2003 p8
The UMBC GMPLS Optical Testbed
l 9 optical nodes– IP/MPLS router– Ethernet switch– OXC
l 4 operational wavelengths– 100/1000Mbps line rate
lWeb-accessible GUI interface
L05-3
5 6 7 811 12 13 14
9 10 19 20
23 32 36 37
P5 P6 P7j5 P8j6 j7 j8
L07-3 L07-4 L08-3 L08-4
5 61211 13 14
22 26
41 42 43 101
P5 P6 P7j5 P8j6 j7 j8
28
L4-4
45
L06-3
15
L10-4
1
L10-3
5
915
103
i101
i2 02
i3o3
i4o4
o3i3
p0
i2o2
i6
i7
o6
o7
i8
o8
i1o1
NIC
1
i5 o5
i4
L9-4
6
j0 p0
L12-2
111
L06-4
16
L12-4
25
L12-1
110
NIC
1N
IC2
NIC
3N
IC4
NIC
5
NIC
1
NIC
2N
IC1
NIC
2
NIC
1
NIC
2
NIC
1N
IC2
NIC
3
NIC
4N
IC5
NIC
2
NIC
2
NIC
1L11-3
31
L12-5
NIC
1
NIC
2
i4 o4
L05-4
38
o4
Category 5 cable
SC connector
Multimode fiber
BNC-SMA cable
Single-mode fiber
PC/SC connectorL0x-3Laser
Transponder
ConverterNICGigabit NIC
GMPLS router
layer2 switch
OXC
ER1-2
CR1 CR2
CR3
ER1-1
ER2-1
ER2-2
ER3-1 ER3-2
L09-3
102
j0
NIC
1
NIC
2
i5o5
i6 o6 i7
i8
o7
o8
i2
o2
i1 o1
o3
i3
o8i8
i7o7
i6
i5
o6
o5
L3-3 L3-4 L04-3
GUI-Server
L12-3
24
Signaling channel
UMBCLTS Review 2003 p9
Traffic Generator and Testing Software
l Software: based on the 3rd party software as follows – Netperf
» Used to generate background traffic» Used to randomly generate traffic» Used to generate multiple user sessions
– Iperf 1.1.1» Used to test the packet drop rate (PDR), Packet Jitter,
User throughput
– ICMP Ping» Use to get the path round-trip latency
UMBCLTS Review 2003 p10
PC throughput capability
93%(iperf)3.52G(iperf)UDP STREAMClientDell SC600
83%2.33GUDP STREAMCR
UDP STREAM
TCP STREAM
TCP STREAM
TCP STREAM
DATA TYPE
70%2.53MER3-1,ER3-2ER2-2,ER2-1
96%(netperf)91~94%(iperf)
2.231G(netperf)2.2G(iperf)
ER3-1,ER3-2ER2-2,ER2-1
98%(netperf)92~96%(iperf)
2.290G(netperf)2.2G(iperf)
CR
95%(iperf)95%(netperf)
3.51G(iperf)4.1G(netperf)
Client Dell SC600
CPU utility rate-system
Maximum TCP (loopback) /UDPthroughput
l Experiment setup: local loopback
UMBCLTS Review 2003 p11
Path Latency Reduction
Note:•Test results are extracted from the 9-node UMBC Optical testbed.•Results only show single node latency.•Equipments used in the measurement : Cisco Catalyst 3550 router, Intel 100Mbps Ethernet switch,
PC-based NIST MPLS switch (Pentium 4 @ 2.0Ghz). •In this test, Layer 2.5 latency is limited by our home-built MPLS switch. Layer 2.5 latency should be
smaller if commercial MPLS switches are used.•Real values of throughput and latency may vary, if different router/switch/OXC devices were used.
However similar performance behavior is expected.
0.10
0.15
0.20
0.25
0.30
100 200 300 400 500 600 700
Layer 3 latencyLayer 2.5 latencyLayer 2 latencyLayer 1 latency
Traffic Load (Mbps)
Late
ncy
(ms)
UMBCLTS Review 2003 p12
Analysis on Bandwidth Utilization Efficiency
Note:•The comparison is based on the 30-nodes NSFNET topology assuming that each node has full L2 O/E/O
capability•Each individual bandwidth request <=2Mbps•All traffics are Internet best-effort traffic•All-optical : adaptive routing and first-fit (FF) wavelength assignment•O/E/O: standard SPF calculation
ã Comparison of system blocking probability ã Comparison of bandwidth utilization on a single node
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
2 4 6 8 10 12 14 16 18 20
All-optical RouteL2 O/E/O Route
Wavelength # = 16
ErLang
Blo
ckin
g P
roba
bilit
y
0.2
0.3
0.4
0.5
0.6
0.7
0 5 10 15 20
L2 O/E/O RouteAll-optical Route
Wavelength # = 16
ErLang
Ban
dwid
th u
tiliz
atio
n
UMBCLTS Review 2003 p13
Efficiency of Traffic Grooming
Note:•Test results are acquired from the UMBC Optical testbed•Equipments used in the measurement : Cisco Catalyst 3550 router, Intel 100Mbps Ethernet switch•Real values of throughput and latency may vary, if different router/switch/OXC devices were used.
However similar performance behavior is expected.
0.6
0.8
1.0
1.2
1 2 3 4 5 6 7 8
Grooming effiency on a 100Mbps Router portGrooming effiency on a 100Mbps Ethernet port
Number of users
Thro
ughp
ut (
100%
)
UMBCLTS Review 2003 p14
GMPLS Operating System Design Issues
for UMBC MPλS Testbed
UMBCLTS Review 2003 p15
The Operating System Is Based On…
l RFC GMPLS, MPLS Draftsl FreeBSD Rel. 3.3l NIST MPLS Switch enginel Alternative Queuing (ALTQ)l ReSerVation Protocol (RSVP) + TE
extensionl OSPF-TE
UMBCLTS Review 2003 p16
Our Special Considerations
l We assume full network-awareness at each node– A cluster network environment
l The QoS condition is treated as a binary constraint– Bandwidth, latency, packet drop rate
l No splitting trafficl Traffic history is not considered
» Routing decisions are made independent of past traffic pattern
l Dynamic cost factor modell Routing calculations based on heuristics
UMBCLTS Review 2003 p17
Routing & Signaling Process
Newrequest
PreferredRouting
Best-effortRouting
Signaling(Path Setup)
DataTransmission
NetworkStatusupdate
Trafficmeasurement
Error noderesponsible to find an
alternative routebetween itself anddestination node
Reject (Pathblocked)
Send errorinformation to
source node, wherean alternative route
is searched
Successful?
PremiumRouting
Successful?
No, Premium/ Preferred
Yes
Reject (Pathblocked)
No,Best-effort
No,Maximumretry timeexceeded
Reject (Pathblocked)
No, Maximum retrytime exceeded
NoNo
No
Successful
No
Yes
Hold request andWait for globalsynchronization
UMBCLTS Review 2003 p18
Routing – Dynamic Problem
l Given – Current network status– A set of QoS requests arriving one at a time
l Objective– Maximize total number of successful connections
l Constraints– Bounded QoS for premium and preferred services– Wavelength continuity constraint at each fiber– Grooming constraint : total bandwidth of premium and
preferred service at any Ethernet switch port cannot exceed the maximum physical outgoing bandwidth
» Assuming the Ethernet switch has DiffServ capability– Reconfiguration policy
UMBCLTS Review 2003 p19
Routing Design
Delay-Constrained Least-Cost Routing (DCLC)
l Graph Extraction Problem l S2 : Design of the Cost Functionl S3 : Design of the Delay Functionl S4 : Routing Heuristic
UMBCLTS Review 2003 p20
Graph Extraction
– Physical node extraction
oeocV ,
AdjacentNode 1
oxcincV ,
AdjacentNode 2
AdjacentNode 3
AdjacentNode 2
AdjacentNode 3
oxcoutcV ,
Wavelength 1
AdjacentNode 1
oxcincV ,
AdjacentNode 2
AdjacentNode 1
AdjacentNode 2
AdjacentNode 3
oxcoutcV ,AdjacentNode 3
Wavelength 2
UMBCLTS Review 2003 p21
Routing – Sub-Problem 1
l S1 : Two-Pass Dijkstra for DCLC– First pass : calculate least delay LDu from
every node u to destination node d and establish the feasible subset of nodes D(LDu’) < D
– Second pass : calculate least cost path from source s to d ; at each relaxation, count only nodes u’ that have D(LDu’) + delay so far <= D
– Time complexity : O(nlog(n)+m)– The path found by DCLC is a feasible path
UMBCLTS Review 2003 p22
Two-Pass Dijkstra : Simulations• NSFNET topology (32 nodes)• Uniformly generated cost&delay on the graph
1
3
2
45
6 17
13
7
9
16
14
15
12
11
10
18
21
1924
2223
25
29
27
28
30
32
31
20
268
UMBCLTS Review 2003 p23
Two-Pass Dijkstra : Simulations• NSFNET topology (32 nodes)• Uniformly distributed cost&delay factors on the graph
200
240
280
320
360
150 200 250 300
Optimal PathLeast Delay PathTwo-Pass Dijkstra
Delay Constraint
Cos
t
.1-1% inefficiency
UMBCLTS Review 2003 p24
Two-Pass Dijkstra : Simulations• NSFNET topology (32 nodes)• Uniformly generated cost&delay on the graph
226.3226.4226.5226.6226.7226.8226.9227.0227.1227.2227.3227.4227.5227.6227.7227.8227.9228.0228.1228.2228.3228.4228.5228.6
225 250 275
Optimal PathLeast Delay PathTwo-Pass Dijkstra
Delay Constraint
Cos
t
UMBCLTS Review 2003 p25
Conclusionsl A novel hybrid network scheme that applies GMPLS on
optical Ethernet is demonstrated– An integrated reconfigurable Ethernet (Layer 2) + Optical layer– QoS traffic engineering
» Efficient bandwidth utilization» Cost-effective
l The testbed is utilized to develop network control and management schemes
– Study advanced control plan issues
l Understand the issues related to actively managed network
– When, where and how to trigger network reconfiguration
l In our system, each data packet connection can be treatedas an independent circuit connection
– The system can be used to study networks with much larger dimension