A Performance Study of IP and MPLS
Traffic Engineering Techniques under
Traffic Variations
Sukrit Dasgupta Department of ECE
Drexel University Philadelphia, PA, USA
Jaudelice C. de Oliveira Department of ECE
Drexel University Philadelphia, PA, USA
[email protected] Jean-Philippe Vasseur Cisco Systems Boxborough, MA, USA [email protected] Presented at IEEE-GLOBECOM PMQRS 09 Washington DC, USA 29th November 2007
Outline
•
Introduction
•
Traffic Engineering
•
Traffic Engineering with MPLS and IP
•
Performance Comparison
Introduction
•
The Internet is more widely used than ever ...
• New applications generating enormous volumes of traffic• Applications such as Gaming, Video On Demand, VoIP, etc., have
strict QoS requirements
• With explosive growth and strict QoS requirements comes the need for
efficient resource management
•
Service Providers resort to ...
• Network Engineering• Manipulate network to suit traffic
• Buy new equipment (fiber, routers, etc.) to keep up with growth
• At 60-70% annual traffic growth rate (200%+ in Japan), proves to be very
expensive and time consuming
• Traffic Engineering
• Manipulate traffic to suit network
• Move traffic in the network to create more space
• Commonly deployed using IP, MultiProtocol Label Swticthing (MPLS) and
Traffic Engineering
•
“Art” of efficiently routing traffic to ...
• Improve efficiency of bandwidth resources • Ensure desirable path for most/all traffic • Reduce operational costs•
Challenges ...
• Current mechanisms require the knowledge of Traffic Matrix
• Mechanisms can be traffic disruptive and unable to cope with rapid
changes/multiple failures, etc.
• Multi-constraint objective functions are needed •
Several models exist ...
• Centralized: Efficient in solving multi-constraint problems but scale
poorly and require multiple cycles of computation and deployment
• Distributed: Highly scalable, dynamic but may be complex to analyze
Traffic Engineering with MPLS and IP
•
Traffic Engineering can be performed using ...
• IP with metric optimization• Have to know traffic matrix, effective when conditions do not change
• MPLS Traffic Engineering •
Constraints
• Do not exceed link capacity (or a fraction of link capacity) • Additional constraints (such as delay, etc.)
•
When traffic changes: Flash crowd / Failures / Misconfigs
• With IP:• Change link metric (B. Fortz: Reoptimizing OSPF/ISIS Weights)
• Triggers Shortest Path Tree computation (O.Bonaventure: FIB Ordering) • Reroutes traffic on ‘new’ shortest path.
• With MPLS:
• Compute Constrained Shortest Path (CSPF: Prune links + SPF)
• Setup new reservation and tear down old one (“Make before break”) • Forward traffic on new path.
Traffic Engineering with MPLS and IP
•
With IP ...
• Global impact: Affects every router, SPT
computation
• Micro-Loops: Traffic in transit is sent back and
forth
• Lack of granularity: Specific flows cannot be
selected
• Slow reactive process: Offline optimization
problem
• On linkup: Change the metric to its old value
X 45 Mbps 40 Mbps to D before link failure DS3: OC3: 140 Mbps to E 155 Mbps Router A Router B Router H Router G Router C Router E Router F Router D Router I 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 3/DS3 3/DS3 1/OC48 1/OC48 1/OC48 B 9 D Cost Next Node G E 6 Cost Next Node Packet Drops 40 Mbps to D after link failure X Router A Router B Router H Router G Router C Router E Router F Tunnel1 10 D Cost Next Node Router D Router I 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 2/OC3 3/DS3 3/DS3 1/OC48 1/OC48 1/OC48 G E 6 Cost Next Node 45 Mbps 40 Mbps to D before link failure DS3: OC3: 140 Mbps to E 155 Mbps 40 Mbps to D after link failure •
With MPLS ...
• Tunnel is created• Path computed takes into account current
network state (dynamic path option)
• Path can also be assigned manually (explicit path
option)
• Reduces chances of congestion
• On linkup, reoptimization can be enabled and the
Performance Comparison
•
Simulation Setup
• Realistic Traffic Profile (Daily variation)
• 4 service provider topologies (OSPF-TE as IGP)
• All flows on the shortest path (Same starting condition) • Link failures (to create a heavy load and traffic shifts)
• Independently ~ U(0,60) minutes • Restored ~ U(0,15) minutes
• Traffic rerouted on link failure (path computation)
•
Performance Metrics
• Link Utilization(how good is IP at handling traffic changes/shifts)
• Path Quality: Ratio of current cost to shortest path cost
(how far is MPLS from the shortest path)
20 30 40 50 60 70 80 90 100 0 200 400 600 800 1000 1200 1400 Bandwidth (Kbps) Time (Minutes) Traffic Profile Plot - Data Traffic
Link Utilization
•
With IP ...
• Link utilization crosses 100% several times • Signifies congestion and packet drops
•
With MPLS
• Path computation after pruning links • CSPF computes paths that can fit traffic
0 50 100 150 200 250 300 Maximum Link Utilization
IP+Failures 0 2000 4000 6000 8000 10000 Time 0 20 40 60 80 100 120 Number of Links 0 50 100 150 200 250 300 Maximum Utilization 0 10 20 30 40 50 60 70 80 90 Maximum Link Utilization
Static TE+Failures 0 2000 4000 6000 8000 10000 Time 0 20 40 60 80 100 120 Number of Links 0 10 20 30 40 50 60 70 80 90 Maximum Utilization
IP
MPLS
Hotspots with IP
Topological View: Max Link Util.
MPLS + Failures
Path Quality with MPLS
•
Topology dependent
• ISP1, SYM have almost all TE-LSPs/Traffic on shortest path
• ISP2 and MESH have 90% TE-LSPs/Traffic close to shortest path • ‘Fatter’ TE-LSPs are on longer paths, need more space
• Priorities can be used for alignment of traffic on shorter paths
1 1.5 2 2.5 3 3.5 4 95 96 97 98 99 100
TE:IGP Path cost ratio
Percentage of TE-LSPs
Distribution of TE:IGP path cost ratio accross TE-LSPs Primary=2543,NHop=0,NNHop=0 MESH SYM ISP1 ISP2 1 1.5 2 2.5 3 3.5 4 90 92 94 96 98 100
TE:IGP Path cost ratio
Percentage of total traffic
Distribution of TE:IGP path cost ratio accross traffic Primary=2543,NHop=0,NNHop=0 MESH
SYM ISP1 ISP2
Concluding ...
•
Summary
• Quantified metrics that capture and allow comparison of MPLS and IP
performance
• Showed that MPLS can help to reduce congestion without any metric
re-computation
• Showed that MPLS can keep more traffic and TE-LSPs close to their
shortest path
•
Contributions
• Compared MPLS and IP performance under similar scenarios • Quantified metrics to motivate the use of MPLS
• Time varying distribution of link utilization to capture congestion instances • Path quality with MPLS