Transcript A Practical Approach for Providing QoS

MPLS
Maximizing the Performance and
Profitability of Optical and Data Networks
July 12 - 13, 2001
Dublin, Ireland
A Practical Approach for
Providing QoS: MPLS and DiffServ
Thomas Telkamp
Director Data Architecture and Technology
Global Crossing Telecommunications, Inc.
[email protected]
Agenda
• Global Crossing MPLS deployment
• Quality of Service?
• A Practical Approach
 Network Design and Capacity Planning
 Differentiated Services
 MPLS for Traffic Engineering
 Fast Reroute and Per-Class TE
 Queuing and Scheduling
• Conclusion
Global Crossing IP Backbone Network
AC-1
North American Crossing
AC-2
PEC
PC-1
MAC
EAC
PAC
SAC
Based on announced network
GBLX MPLS Deployment
• Operational since 2Q 1999
 Traffic Engineering
 IP TTL issues
• Worldwide MPLS mesh 1Q 2001
• Currently over 6000 LSPs
• Network:
 Cisco and Juniper routers
 OC-48 wavelengths
 Covering Asia, US, South America and Europe
• New Services:VPN (L2/L3)
MPLS Traffic Engineering
Quality of Service?
Based on a paper with XiPeng Xiao (Photoris, Inc.)
and Lionel M. Ni (Michigan State University)
• Best Effort (e.g. Internet)
• Real-time/Mission-critical traffic (e.g.Voice)
• Increase revenue by value-added services
• Two extremes:
 Overprovisioning of bandwidth without additional
mechanisms
 Sophisticated mechanisms such as per-flow
classification/policing/queuing and scheduling
What Causes Problems?
• Overloaded servers, or access to servers
 Web, E-mail, etc.
• TCP stack implementations
• Link failures
 fiber cuts
 transmissions equipment failures
• Router failures
 complex software
 early deployment of features
 configuration
A Practical Approach
• Good Network Design
• Differentiated Services (DiffServ)
• Traffic Engineering
• Traffic Protection (Fast Reroute)
• Class-based Queuing
• Not:
 Extremely complex schemes (e.g. per-flow)
• affecting equipment reliability
• difficult to configure and manage
Network Design
• Avoid single points of failure
• No bottlenecks in normal condition
• Overprovisioning
 with use of TE network can handle all traffic, even
when the most critical links fails
• Routing (IGP and BGP)
• Security and Denial of Service attacks
• Capacity Planning
Differentiated Services
• How many classes?
 What are the targeted applications for each class?
 Can end users distinguish between classes?
• Example:
 Class 1: Real-time
• application: voice
 Class 2: Assured
• application: trading, non-interactive audio and video
 Class 3: Best Effort
• application: Internet
MPLS Deployment
• Traffic Engineering
 Avoid congestion caused by uneven traffic
distribution
 Macro control
 Constraint based LSP setup
• Two LSP meshes:
 Real-time traffic vs Assured/Best Effort
 Classification based on interface or multi-field lookup
 Different metrics
• LSP Hierarchy
 Scalability and VPNs
MPLS LSP Deployment
Traffic Protection
• IGP convergence (OSPF/IS-IS) takes seconds
 But can be improved by timer and SPF tuning
• see Packet Design paper
• MPLS Fast Reroute
 Link or Node protection
 Pre-configured patch LSPs (sub-optimal)
 Use for real-time traffic only, or for all traffic (based
on implementation)
MPLS Fast Reroute
• Protecting router switches traffic to preconfigured patch LSP after failure detection (fast)
• Ingress router reroutes LSP (slow)
MPLS Cloud
(2) patch LSP activated
(1) failure detected
Ingress
protected segment
Protecting
LSR(k)
router
(3) lsp rerouted
Egress
Per-Class Traffic Engineering
• Avoid concentration of real-time traffic at any link
• Set upper limit on bandwidth reservations per
class
 E.g. max. 40% of a link for VoIP traffic
• IETF Internet Draft(s) on ‘Diff-Serv-aware MPLS
Traffic Engineering’ (Francois Le Faucheur, et al.)
Class-based Queuing
• Prefer ‘higher’ classes during congestion
 sub-optimal fast-reroute period
 major failures
• Different queuing/scheduling mechanisms
 Strict Priority Queuing
• Jitter control for EF traffic
 WRR/WFQ
 and combinations
• Configuration issues...
Random Early Detection
• Buffer Management
 prevent tail-drop
• TCP oscillations and synchronization
 RED drops based on average queue length
 WRED drops with different probability for each
class
• Only during congestion
• Not used to guarantee bandwidth!
Conclusion
• Use combination of good network design, overprovisioning and MPLS/DiffServ
• Use Traffic Engineering to prevent congestion
• Use fast reroute and priority queuing for real-time
traffic
• Use WRR/WFQ to differentiate between Assured
and Best Effort traffic
• Too complex and too many features will make the
network unreliable/unstable
Questions?