You are on page 1of 90

Quality of Service and Traffic Engineering using Multiprotocol Label Switching

Raj Jain Professor of Computer and Information Sciences The Ohio State University Columbus, OH 43210-1277 jain@cis.ohio-state.edu These slides are available on-line at: http://www.cis.ohio-state.edu/~jain/talks/mpls_te.htm
The Ohio State University

Raj Jain 1

Overview
1. MPLS Overview 2. Traffic Engineering using MPLS 3. Our Simulation Results 4. Other QoS Approaches and their Interoperability with MPLS
The Ohio State University

Raj Jain 2

Part 1: MPLS Overview
q q q q q

Routing vs Switching Label Switching Concepts Label Stacks Label Distribution Protocol Independent vs Ordered Control

The Ohio State University

Raj Jain 3

Max prefix match. SONET Raj Jain 4 The Ohio State University .Routing vs Switching 164. Ethernet. ⇒ Search Operation ⇒ Complexity ≈ O(log2n) Switching: Based on circuit numbers ⇒ Indexing operation ⇒ Complexity O(1) ⇒ Fast and Scalable for large networks and large address spaces These distinctions apply on all datalinks: ATM.107.201 q 3 q q Routing: Based on address lookup.61.

Routing vs Switching over ATM Router ATM Host Switch ATM Host On ATM networks: q IP routers use IP addresses ⇒ Reassemble IP datagrams from cells q IP Switches use ATM Virtual circuit numbers ⇒ Switch cells ⇒ Do not need to reassemble IP datagrams ⇒ Fast The Ohio State University 5 Raj Jain .

IP has no traffic engineering ⇒ Under/over-utilized links Logical ≠ Physical ⇒ ATM has n2 scaling problem MPLS takes the best of both IP and ATM networks Works on both ATM and non-ATM networks ⇒ Easier management Raj Jain 6 The Ohio State University .High-Speed Backbone Alternatives q q q q q Physical Topology Logical Topology High-speed (OC-3 and higher) ATM switches easily available. IP routers either not available or expensive.

Unlabeled Labeled R H H R H R R H H The Ohio State University Raj Jain 7 . Exit router strips it off. Do not need to look inside ⇒ Fast.Label Switching q q q Label = Circuit number = VC Id Ingress router/host puts a label. Switches switch packets based on labels.

Label Switching (Cont) q q Labels have local significance Labels are changed at every hop 1 164.* 2 2 2 128.146.* 1 3 The Ohio State University Raj Jain 8 .61.146.107.7.210 Input Input Adr Output Output Port Label Prefix Port Label 1 1 164.146.* 164.61.61.*.107.48 1 1 128.107.146.* 2 2 2 2 128.48 2 164.7.210 3 128.107.61.*.

.. Not specific to a routing protocol (OSPF.. ATM or LANs.) The Ohio State University Raj Jain 9 . RIP. Not specific to ATM. Technology extendible to other L3 protocols.MPLS q q q q q Multiprotocol Label Switching IETF working group to develop switched IP forwarding Initially focused on IPv4 and IPv6.

locally significant Label Switching Router (LSR): Routers that use labels Forwarding Equivalence Class (FEC): Same Path + treatment ⇒ Same Label MPLS Domain: Contiguous set of MPLS nodes in one Administrative domain MPLS edge node = Egress or ingress node Label distribution protocol ≅ Routing protocols MPLS Domain The Ohio State University Raj Jain 10 .MPLS Terminology q q q q q q Label = Short fixed length. physically contiguous.

.. 2=IPv6) L2 Header Label 1 Label 2 . Label n 2 The Ohio State University 2 2 1 1 11 1 2 2 2 Raj Jain .Label Stacks q q Label q q A MPLS packet may have multiple labels Labels are pushed/popped as they enter/leave MPLS domain Stack allows hierarchy of MPLS domains Bottom label may indicate protocol (0=IPv4.

Private Net A Private Net B The Ohio State University Public/ISP Net 12 Private Net B Private Net A Raj Jain . Net A and B can use the same private addresses. BGP/OSPF Routing Hierarchy B B BO O OB B B 2.Label Stack Examples 1. VPN: Top label used in public network.

. 1⇒ Bottom of Stack L2 Header Label Stack Entry Label Stack Entry .Label Stack Entry Format q q q q Labels = Explicit or implicit L2 header TTL = Time to live Exp = Experimental SI = Stack indicator.. 20b 3b 1b 8b Label Exp SI TTL The Ohio State University Raj Jain 13 .

RSVP is being extended to allow label request and response Raj Jain 14 The Ohio State University .. Many existing routing protocols are being extended: BGP. when a packet arrives ⇒ latency A new Label Distribution Protocol called LDP is being defined.Label Assignment q q q q Unsolicited: Topology driven ⇒ Routing protocols exchange labels with routing information. e.g. OSPF On-Demand: ⇒ Label assigned when requested.

q Session establishment messages sent on TCP port 646. All others TCP. The Ohio State University Raj Jain q 15 . q Hello messages are sent on UDP port 646. multipath. or QoS in the first version. q No multicast. q LDP messages: m Session establishment/termination messages m Discovery messages to announce LSRs (Hello) m Advertisement msgs to create/delete/change label m Notification messages for errors and advice q Discovery messages are UDP based.Label Distribution Protocol LDP peers: LSRs that exchange LDP messages. Using an LDP session.

q q q q q q q q q q q q q Hello Initialization Label Request Label Mapping (Label Response) Label Withdraw (No longer recognized by downstream) Label Release (No longer needed by upstream) Label Abort Request KeepAlive Notification Address (advertise interface addresses) Address Withdraw Vendor-Private Experimental Raj Jain 16 LDP Messages The Ohio State University .

LDP TLVs q q q q q q q q q q q q FEC (Wild card. or host address) Address List Hop Count Path Vector Generic Label ATM Label Frame Relay Label Status Extended Status Returned PDU Returned Message Common Hello parameters Raj Jain 17 The Ohio State University . prefix.

May cause loops. Ordered: A router issues labels for an FEC only if it is the egress router or if it has received a label from the next hop ⇒ Use LSP only after it is fully setup Use ordered LSP control if you need QoS for LSP LSRs can use either method.Independent vs Ordered Control q q q q Independent: Each router issues Labels for FECs. Optional The Ohio State University Raj Jain 18 .

q No TTL field ⇒ Hops between ingress and egress can be computed during LSP setup. Ingress router drops if TTL < hops to egress q ATM LSRs need to participate in network layer routing protocols (OSPF. VPI All ATM switches should use the same encoding. q No Stack bit ⇒ Maximum two possible levels of hierarchy: VCI.MPLS Over ATM With MPLS software. BGP) q VPI/VCI space may be segmented for label switching and normal ATM switching The Ohio State University Raj Jain q 19 . ATM switches can act as LSRs. q VPI/VCI fields are used for labels.

Stream Merging q q q q Required for egress based labels. VP Merge: VPI = Labels. Delay. VCI = source 5 5 LAN Switch 5 5 5 3 3 20 5 5 5 ATM Switch 3 3 3 Raj Jain The Ohio State University . Helpful for mpt-topt streams. VC-merge: Store all cells of a frame and forward together ⇒ Need more buffering. In ATM/AAL5. cells of frames on the same VC cannot be intermingled ⇒ VCs cannot be merged.

Works on all media: ATM and non-ATM. q LDP allows independent or ordered control The Ohio State University q 21 Raj Jain . q Common routing protocols and RSVP are being extended to include label exchange. q Label is similar to circuit number or VC Id. q Label stacks allow hierarchy of MPLS domains.Summary of Part 1: MPLS MPLS combines the best of ATM and IP.

Part 2: Traffic Engineering q q q q q q Objectives and Mechanisms Traffic Trunks CR-LDP Explicit Route Priority and Preemption Traffic Engineering Extensions to OSPF and IS-IS The Ohio State University Raj Jain 22 .

g. min loss.Traffic Engineering Objectives q q q q q User’s Performance Optimization ⇒ Maximum throughput. min delay variation Efficient resource allocation for the provider ⇒ Efficient Utilization of all links ⇒ Load Balancing on parallel paths ⇒ Minimize buffer utilization m Current routing protocols (e.. QoS Guarantee: Selecting paths that can meet QoS Enforce Service Level agreements Enforce policies: Constraint based routing ⊇ QoSR Raj Jain 23 The Ohio State University . RIP and OSPF) find the shortest path (may be over-utilized). Min delay.

Traffic Engineering Components Signaling 1 and Admission control 2 Shaping 3 Policing Scheduling 5 4 Routing 7 Traffic Monitoring 6 and feedback The Ohio State University Buffer Mgmt Raj Jain 24 .

Shaping: Smoothen the bursts 3. Admission Control: Network may deny the request. Re-optimization/Pinning.Traffic Engineering Components 1. Explicit Accounting/Billing Performance Monitoring/Capacity Planning The Ohio State University Raj Jain 25 . Policing: Ensure that users are following rules 4. Prioritization. Request Prioritization. Preemption. Preemption 6. Signaling: Tell the network about traffic and QoS. Routing: Path Selection. Drop Priority 7. Fault Recovery 5. Buffer Management: Drop Thresholds. Feedback: Implicit. 2. Scheduling: Weight.

Routing Explicit routing of LSPs Constrained based routing of LSPs Allows both Traffic constraints and Resource Constraints (Resource Attributes) Hierarchical division of the problem (Label Stacks) Traffic trunks allow aggregation and disaggregation (Shortest path routing allows only aggregation) The Ohio State University Raj Jain 26 .MPLS Mechanisms for TE q q q q q Signaling. Admission Control.

Raj Jain 27 The Ohio State University . LSP ⇒ Next hop Class can be coded in Exp or Label field. Assume Exp.Traffic Trunks q q q Trunk: Aggregation of flows of same class on same LSP Trunks are routable ⇒ LSP through which trunk passes can be changed Class ⇒ Queue.

Trunks vs LSPs Tour Group First Class Business Class Flights = LSP Tour Groups = Trunks The Ohio State University Coach Class Raj Jain 28 .

and Links q q q Label Switched Path (LSP): Path for all packets with the same label Trunk: Same Label+Exp Flow: Same MPLS+IP+TCP headers DL Label Exp SI TTL IP TCP Flows Trunk LSP LSP The Ohio State University Link Raj Jain 29 . Trunks.Flows. LSPs.

Trunk paths are setup based on policies or specified resource availability.Traffic Trunks q q q q q Each traffic trunk can have a set of associated characteristics. priority. policing Some trunks may preempt other trunks. or nonpreemptable. preemptable.g. A trunk can be preemptor. preemption.. Multiple LSPs can be used in parallel to the same egress. non-preemptor. e. 30 The Ohio State University Raj Jain . Trunks can be rerouted. A traffic trunk can have alternate sets of paths in case of failure of the main path.

Fault Recovery Scheduling: Class Weight. Re-optimization/Pinning.Trunk Attributes q q q q q q q q q q Signaling: Routing Protocols. CR-LDP Admission Control: Network may deny the request. RSVP. Preemption. Prioritization. Policing: Token Bucket Shaping: Smoothen the bursts Routing: Path Selection. Preemption Buffer Management: Class drop thresholds/priority Feedback: Implicit. Explicit (ICMP being discussed) Accounting/Billing Performance Monitoring/Capacity Planning Raj Jain 31 The Ohio State University . Request Prioritization.

Leaky bucket sends less than rt. On bursty arrivals after a long idle: m Token bucket results in bursty departures m Leaky bucket results in smooth departures Raj Jain 32 The Ohio State University .Token Bucket vs Leaky Bucket Constant Rate Tokens Bursty Arrivals Bursty Arrivals Tokens? Yes No Bucket no full? Yes q q q Both designed for controlling average rate. Token bucket sends less than b+rt. Used by ATM. Used by IETF.

Token and Leaky Bucket
Constant Rate Tokens Bursty Arrivals Tokens? yes No
q q

Policing and Shaping Leaky bucket rate = Peak rate > Token Rate = Average rate

Bucket full? Yes
The Ohio State University

no

Smooth Departures
Raj Jain

33

Traffic Granularity
q

q q q

q q

Same label ⇒ Same port quadruples (source/destination address, IP protocol, source/destination port) Same QoS + Port quadruples Same host pair (Source/destination address) Same network pairs (Source/destination address prefixes) Same destination network Same Egress router
Raj Jain 34

The Ohio State University

Traffic Granularity (Cont)
q q q q

Same BGP next hop AS Same BGP destination AS Same Shared multicast tree (*,G) Same Source specific multicast tree (S,G)

The Ohio State University

Raj Jain 35

CR-LDP
q q

Extension of LDP for constraint-based routing (CR) New Features: m Traffic parameters m Explicit Routing m Preemption of existing route. Based on holding priority of existing route and setup priority of new route m Route pinning: To prevent path changes

The Ohio State University

Raj Jain 36

CR-LDP (Cont)
q q

q

q q

No new messages Enhanced Messages: Label request, Label Mapping, Notification New TLVs: Explicit Route, Explicit Route Hop, Traffic, Route Pinning, Resource Class, Pre-emption, LSP Id Enhanced TLVs: FEC (CRLSP) Each setup (label request) message has a unique connection ID (LSPID)
Raj Jain 37

The Ohio State University

CR-LSP Setup q q q q New CR-TLV ⇒ Use “downstream on demand” label advertisement with ordered control Similar to ATM connection setup message. Egress router indicates the negotiated values in the response (label mapping message) Other LSRs return the response towards the ingress and reserve. The Ohio State University Raj Jain 38 .

committed data rate. committed burst size.Traffic Parameters q q q q q The Ohio State University DP0 DP1 DP2 Single-rate dual-token-bucket Tokens generated at “Committed Data Rate” (CDR). if full go to the 2nd bucket Peak. excess burst size (Dual-bucket single rate) Negotiation Allowed Color Aware ⇒ Use incoming drop precedence (DP) Color unaware ⇒ Ignore incoming drop precedence 39 Raj Jain . Tokens go to 1st bucket.

MPLS tunnels or Autonomous systems Example: R1-R2-Net B-R7-R8 R3 Net B R4 R7 Net D R8 R5 Net C R6 Raj Jain 40 R1 Net A R2 The Ohio State University . IPv6 prefix.Explicit Route q q q Explicit route specified as a list of Explicit Route Hops (group of nodes) Hops can include IPv4 prefix.

Explicit Route (Cont) q q q All or a subset may be traversed The list is specified by edge router based on imperfect info (Strict/loose) m Strict ⇒ Path must include only nodes from the previous and this abstract node m Loose ⇒ path between two nodes may include other nodes Managed like ATM PNNI Designated Transit Lists (DTLs) Raj Jain 41 The Ohio State University .

Path Selection q q q q q q Manual/Administrative Dynamically computed Explicitly specified: Partially/fully. Single/Set Non-Mandatory ⇒ Use any available path if specified not available Set ⇒ Preference ordered list Resource class affinity The Ohio State University Raj Jain 42 . strict/loose. Mandatory/non-mandatory.

transit/local-only A resource can be member of multiple classes The Ohio State University Raj Jain 43 .Resource Attributes q q q q q Capacity Overbooking Factor: Maximum Allocation Multiplier Class: Allows policy enforcement Class Examples: secure/non-secure.

Resource Class Affinity q q q q Each resource has a class Affinity = Desirability Binary Affinity: 0 ⇒ Must Exclude. Not-specified ⇒ Don't care <Class. 1 ⇒ Must Include. affinity> pair can be used to implement policies The Ohio State University Raj Jain 44 .

Resilience /⇒ Adaptivity Idea: Adaptivity is not binary ⇒ Rerouting period The Ohio State University Raj Jain 45 .Adaptivity and Resilience q q q q q q q Stability: Route pinning Resource availability is dynamic Trunks can live for long time Adaptivity: Re-optimization when availability changes Resilience: Reroute if path breaks Adaptivity ⇒ Resilience.

Priority and Preemption q q q q q Preemptor-enabled: Can preempt other trunks Non-Preemptor: Can't preempt other trunks Preemptable: Can be preempted by other trunks Non-Preemptable: Can't be preempted by other trunks These attributes and priority are used to decide preemption The Ohio State University Raj Jain 46 .

Traffic Engineering Extensions to OSPF q q q q q q q Add to Link State Advertisements: TE Metric: May be different from standard OSPF link metric Maximum bandwidth Maximum Reservable Bandwidth: May be more than maximum bandwidth Unreserved Bandwidth Resource Class/color Ref: draft-katz-yeung-ospf-traffic-00.txt Raj Jain 47 The Ohio State University .

TE Extensions to OSPF (Cont) q q Link Delay and Link Loss rate also proposed in draftwimer-ospf-traffic-00. TE tunnels are used as links to tunnel egress ⇒ LSP The Ohio State University Raj Jain 48 .txt In path calculations.

Traffic Engineering Extensions to IS-IS q q q q q q q Add to Link State Protocol Data Units: TE Metric Maximum bandwidth Maximum Reservable Bandwidth: May be more than maximum bandwidth Unreserved Bandwidth Resource Class/color Ref: draft-ietf-isis-traffic-01.txt Raj Jain 49 The Ohio State University .

Summary of Part 2: Traffic Engg q q q q Goal of traffic engineering is to optimize performance for users and providers and ensure QoS MPLS traffic trunks are like ATM VCs that can be routed based on explicit route or policies CR-LDP allows explicit routing. and QoS OSPF and IS-IS is being modified for traffic engg Raj Jain 50 The Ohio State University . traffic parameters. constraint-based routing.

UDP m Case 4: Non End-to-End Trunks Future Work Raj Jain 51 The Ohio State University .A Simulation Analysis of Traffic Engineering q q q Simulation Model Four Simulation Scenarios m Case 1: No Trunks. No MPLS m Case 2: Two trunks w UDP + TCP Mixed m Case 3: Three Trunks w Isolated TCP.

UDP MSS = 210B 52 The Ohio State University Raj Jain . TCP2 MSS = 1024 B. All links have 5ms delay TCP1 MSS = 512 B.. R2-R4-R5 is a low bandwidth (15 Mbps) path.Simulation Model Src 1 Src 1 Src n Src n q q q q q R3 60 Mbps 45 Mbps 60 Mbps R1 R2 R5 R6 Dest 1 Dest 1 15 Mbps R4 Dest n Dest n Sources 1.n R2-R3-R5 is a high bandwidth (45 Mbps) path..n send TCP and UDP packets to Dest 1.

Normal IP with Best Effort routing 2. Two trunks using Label Switched Paths m Trunk 1: R1-R2-R3-R5-R6 q TCP and UDP sources are multiplexed over this trunk m Trunk 2: R1-R2-R4-R5-R6 q Only TCP sources over this trunk 3. The Ohio State University Raj Jain 53 .Simulation Scenarios 1. Non End-to-end trunks. Three trunks using Label Switched Paths m All three flows are isolated. 4.

No MPLS 45 40 Throughput (Mbps) 35 30 UDP UDP TCP1 TCP2 25 20 15 TCP2 TCP1 15 20 25 30 35 40 45 10 5 0 UDP Rate (Mbps) q q q 15 Mbps path not used at all TCP suffers as UDP increases its rate Unfairness among TCP flows 54 The Ohio State University Raj Jain .Case 1: No Trunks.

Two trunks w UDP + TCP Mixed 50 45 Throughput (Mbps) 40 35 UDP UDP TCP1 TCP2 30 25 20 15 TCP2 TCP1 15 20 25 30 35 40 45 10 5 0 q q q Total throughput > 45 Mbps (both paths used) TCP flows sharing the trunk with UDP suffer TCP flow not sharing with UDP do not suffer Raj Jain 55 UDP rate (Mbps) The Ohio State University .

3 Trunks w Isolated TCP. UDP 35 30 Throughput (Mbps) UDP TCP1 UDP TCP1 TCP2 25 20 15 10 5 TCP2 15 20 25 30 35 40 45 0 UDP rate (Mbps) q TCP flows are not affected by UDP and achieve a fairly constant throughput Raj Jain 56 The Ohio State University .

Non End-to-End Trunks 45 40 Throughput (Mbps) 35 30 UDP UDP TCP1 TCP2 25 20 15 TCP2 TCP1 15 20 25 30 35 40 45 10 5 0 UDP rate (Mbps) q TCP flows are affected by UDP in the shared path Raj Jain 57 The Ohio State University .

Tail drop Round Trip Time TCP parameters: MSS. DiffServ vs MPLS. WF2Q+ Packet drop policies: RED.Future Work q q q q q q Other Traffic Scenarios: m Aggregate flows: TCP+UDP m Short duration TCP connections m Bursty (Web) traffic Queue Service Policies: WFQ. DiffServ+MPLS Raj Jain 58 The Ohio State University . window size. WF2Q. etc.

Summary of Part 3: TE Analysis q q q Total network throughput improves significantly with proper traffic engineering Congestion-unresponsive flows affect congestionresponsive flows m Separate trunks for different types of flows Trunks should be end-to-end m Trunk + No Trunk = No Trunk Raj Jain 59 The Ohio State University .

Part 4: Other QoS Approaches and MPLS Interoperability q q q q ATM Integrated Services/RSVP Differentiated Services IEEE 802.1D The Ohio State University Raj Jain 60 .

delay variation rt-VBR: Throughput. delay. Feedback. delay variation nrt-VBR: Throughput UBR: No Guarantees GFR: Minimum Throughput ABR: Minimum Throughput.ATM Service Categories q q q q q q q CBR: Throughput. Very low loss. delay. ATM also has QoS-based routing (PNNI) The Ohio State University Raj Jain 61 .

The Ohio State University Raj Jain q 62 .ATM QoS: Issues Can’t easily aggregate QoS: VP = Σ VCs q Can’t easily specify QoS: What is the CDV required for a movie? q Signaling too complex ⇒ Need Lightweight Signaling q Need Heterogeneous Point-to-Multipoint: Variegated VCs q Need QoS Renegotiation q Need Group Address q Need priority or weight among VCs to map DiffServ and 802.1D q MPLS also has many of these problems.

. m Every element along the path must provide delay bound. Shared Ethernet.Integrated Services q q q Best Effort Service: Like UBR. Controlled-Load Service: Performance as good as in an unloaded datagram network. No quantitative assurances. m Like CBR or rt-VBR Raj Jain 63 The Ohio State University . m Is not always implementable.g. m Delay jitter or average delay not guaranteed or minimized. Like nrt-VBR or UBR w MCR Guaranteed Service: rt-VBR m Firm bound on data throughput and delay. e.

RSVP q q q q Resource ReSerVation Protocol Internet signaling protocol Carries resource reservation requests through the network including traffic specs. network resource availability Sets up reservations at each hop Traffic Spec Network Traffic Spec Sender Receiver QoS Spec Available Resources AdSpec The Ohio State University Raj Jain 64 . QoS specs.

Number of flows in the backbone may be large.Problems with IntServ/RSVP q q q q Complexity in routers: packet classification. ⇒ Suitable for small private networks Need a concept of “Virtual Paths” or aggregated flow groups for the backbone Need policy controls: Who can make reservations? Support for accounting and security. Raj Jain 65 The Ohio State University . scheduling Per-Flow State: O(n) ⇒ Not scalable with # of flows. ⇒ RSVP admission policy (rap) working group.

RSVP does not have negotiation and backtracking Throughput and delay guarantees require support of lower layers. Raj Jain 66 The Ohio State University . MPLS solves many of these problems. Need switched full-duplex LANs.Problems (Cont) q q q q q Receiver Based: Need sender control/notifications in some cases. Which receiver pays for shared part of the tree? Soft State: Need route/path pinning (stability). Shared Ethernet ⇒ IP can’t do GS or CLS. Limit number of changes during a session.

MPLS over IntServ: Not a realistic scenario. IntServ over MPLS: m MPLS can provide controlled service.MPLS-IntServ Interoperability MPLS IntServ MPLS q IntServ MPLS IntServ q q IntServ is more complex and will be less widely implemented. guaranteed service and best effort services without the need for classification at each hop. Raj Jain 67 The Ohio State University .

Differentiated Services Ver Hdr Len Precedence ToS Unused Tot Len 4b 4b 3b 4b 1b 16b q IPv4: 3-bit precedence + 4-bit ToS q OSPF and integrated IS-IS can compute paths for each ToS q Many vendors use IP precedence bits but the service varies ⇒ Need a standard ⇒ Differentiated Services q DS working group formed February 1998 q Charter: Define ds byte (IPv4 ToS field) q Mail Archive: http://www-nrg.ee.gov/diff-serv-arch/ The Ohio State University Raj Jain 68 .lbl.

. Meter Packets Classifier Marker 69 Shaper/Dropper Raj Jain The Ohio State University . Markers (tag). DS-byte. Shapers (delay). ports. Droppers (drop) Behavior Aggregate (BA) Classifier: Based on DS byte only Multi-field (MF) Classifiers: Based on IP addresses. etc.DiffServ Concepts q q q q Micro-flow = A single application-to-application flow Traffic Conditioners: Meters (token bucket).

controlled load. Video.... guaranteed service m DS group will not develop services They will standardize “Per-Hop Behaviors” The Ohio State University Raj Jain 70 .. WWW.Diff-Serv Concepts (Cont) q Service: Offered by the protocol layer m Application: Mail. Express Delivery... FTP. m Transport: Delivery. Best effort.

g. e. loss priority. PHBs in the group share common constraints..Per-hop Behaviors In q q q q q PHB Out Externally Observable Forwarding Behavior x% of link bandwidth Minimum x% and fair share of excess bandwidth Priority relative to other PHBs PHB Groups: Related PHBs. relative delay Raj Jain 71 The Ohio State University .

Expedited Forwarding q q q q q q q Also known as “Premium Service” Virtual leased line Similar to CBR Guaranteed minimum service rate Policed: Arrival rate < Minimum Service Rate Not affected by other data PHBs ⇒ Highest data priority (if priority queueing) Code point: 101 110 The Ohio State University Raj Jain 72 .

Assured Forwarding q q q PHB Group Four Classes: No particular ordering Three drop preference per class Raj Jain 73 The Ohio State University .

Assured Forwarding (Cont) q q q DS nodes SHOULD implement all 4 classes and MUST accept all 3 drop preferences. Low Medium High Class 1 010 000 010 010 010 100 Class 2 011 000 011 010 011 100 Class 3 100 000 100 010 100 100 Class 4 101 000 101 010 101 100 Raj Jain 74 q Avoids 11x000 (used for network control) The Ohio State University . Similar to nrt-VBR/ABR/GFR Code Points: Drop Prec. Can implement 2 drop preferences.

q End-to-end ≠ Σ per-Hop Designing end-to-end services with weighted guarantees at individual hops is difficult. q Multicast ⇒ Difficult to provision Dynamic multicast membership ⇒ Dynamic SLAs?Jain The Ohio State University Raj q 75 .Problems with DiffServ per-hop ⇒ Need at every hop One non-DiffServ hop can spoil all QoS This applies to almost all QoS approaches. q Designed for static Service Level Agreements (SLAs) Both the network topology and traffic are highly dynamic. Only EF will work.

DiffServ Problems (Cont) q q q q DiffServ is unidirectional ⇒ No receiver control Modified DS field ⇒ Theft and Denial of service. Ingress node should ensure. Multi-campus enterprises need inter-campus QoS. How to ensure resource availability inside the network? QoS is for the aggregate not per-destination. A B C D Raj Jain 76 The Ohio State University .

Only ISPs.DiffServ Problems (Cont) q q The Ohio State University QoS is for the aggregate not micro-flows. m Long flows (voice and video sessions) need perflow guarantees. Not intended/useful for end users. m High-bandwidth flows (1 Mbps video) need perflow guarantees. All IETF approaches are open loop control ⇒ Drop Closed loop control ⇒ Wait at source Data prefers waiting ⇒ Feedback 77 Raj Jain . m Large number of short flows are better handled by aggregates.

DiffServ Problems (Cont) q Guarantees ⇒ Stability of paths ⇒ Connections (hard or soft) Need route pinning or connections. The Ohio State University Raj Jain 78 .

Raj Jain 79 The Ohio State University . MPLS over DiffServ: No end-to-end guarantees ⇒ Not useful DiffServ over MPLS: m DS byte can be encoded in CR-LDP label requests and responses.MPLS-DiffServ Interoperability MPLS DiffServ MPLS DiffServ MPLS DiffServ q q q MPLS is borrowing the best of DiffServ and can be end-to-end.

1Q header Prot Type Payload FCS q Up to eight priorities: Strict. 1 Background 2 Spare 0 Best Effort 3 Excellent Effort 4 Control load 5 Video (Less than 100 ms latency and jitter) 6 Voice (Less than 10 ms latency and jitter) 7 Network Control The Ohio State University Raj Jain 80 CFI = Canonical Format Indicator (Source Routing) .IEEE 802.1D Model Dest Addr Src Addr Tag Prot ID Pri CFI VLAN ID 802.

802. Raj Jain 81 The Ohio State University . Lower priority traffic from other nodes can get through.MPLS-802.1D Traffic over MPLS: m Packet priority can be encoded in Exp field.1D Interoperability MPLS q 802.1D: Priority among packets at the same node.1D MPLS 802.1D q MPLS over 802.1D MPLS 802. label m Trunk priority can be encoded in CR-LDP label requests and responses.

ATM. Switched LANs/PPP in Stub IntServ/RSVP. MPLS in Stub networks DiffServ. ATM.1D. MPLS in the core Switched LANs/PPP ATM/PPP Switched LANs/PPP IntServ/RSVP.1D. MPLS IntServ/RSVP.End-to-end View q q q ATM/PPP backbone. MPLS R R The Ohio State University R R Edge R Core 82 R R Edge R R Raj Jain .802.802. MPLS DiffServ. 802.1D.

QoS Debate Issues q q q q q q q q q q q q Massive Bandwidth vs Managed Bandwidth Per-Flow vs Aggregate Source-Controlled vs Receiver Controlled Soft State vs Hard State Path based vs Access based Quantitative vs Qualitative Absolute vs Relative End-to-end vs Per-hop Static vs Feedback-based One-way multicast vs n-way multicast Homogeneous multicast vs heterogeneous multicast Single vs multiple bottlenecks: Scheduling Raj Jain 83 The Ohio State University .

3D Massive Both Unicast Source.Comparison of QoS Approaches Issue Massive Bandwidth vs Managed Bandwidth Per-Flow vs Aggregate Source-Controlled vs Receiver Controlled Soft State vs Hard State Path based vs Access based Quantitative vs Qualitative ATM Managed IntServ Managed DiffServ Massive MPLS Managed IEEE 802. Multicast both Hard Path Per-flow Receiver Aggregate Both Ingress Both Aggregate Source Soft Path None Access Hard Path Hard Access Qualitative Quantitativ Quantitativ Mostly Both e e+Qualitati qualitative ve Absolute vs Relative Absolute Absolute Mostly Absolute Relative plus relative The Ohio State University 84 Relative Raj Jain .

Comparison (Cont) Issue End-to-end vs Perhop Static vs Feedbackbased One-way multicast vs n-way multicast Homogeneous multicast vs heterogeneous multicast Single vs multiple bottlenecks: Scheduling ATM e-e Both IntServ e-e Static DiffServ Per-hop Static MPLS e-e Static IEEE 802.3D Per-hop Static Only oneway Homogene Heterogen N/A ous eous Homogene N/A ous Multiple Multiple bottleneck Multiple The Ohio State University Raj Jain 85 .

Summary of Part 4 q q q MPLS is taking the best features of ATM. Raj Jain 86 The Ohio State University .1D QoS approaches ⇒ MPLS is most promising MPLS provides a superset of functionality of many of these other technologies Features ⇒ Complexity Complexity has to be controlled. IntServ. and 802. DiffServ.

edu/~jain/cis788-97/ip_switching/index.htm IP Switching: Issues and Alternatives.cis. http://www. and IP Switching.cis.cis. http://www.ohiostate.htm References on QoS over IP.edu/~jain/talks/ipsw. http://www. MPLS. http://www.edu/~jain/refs/ipqs_ref.htm IP Switching.ohiostate.cis.ohio-state.htm 87 The Ohio State University Raj Jain .edu/~jain/refs/ipsw_ref.ohio-state. http://www.ohiostate.cis.References q q q q q References on MPOA.edu/~jain/papers/mpls-te-anal.htm Quality of Service using Traffic Engineering over MPLS: An Analysis.

cis. draft-ietf-mpls-traffic-eng-01.juniper.html Raj Jain 88 The Ohio State University .fm.txt Constraint-based LSP Setup using LDP.ohio-state. http://www.htm Requirements for Traffic Engineering over MPLS. draft-ietfmpls-cr-ldp-01.References (Cont) q q q q Quality of Service in IP Networks. http://www.edu/~jain/talks/ipqos.net/techcenter/techpapers/optimizi ng-routing-sw.txt Optimizing Routing Software for Reliable Internet Growth.

http://www.com/warp/public/784/packet/apr99/6 .Multiprotocol Label Switching.html The Ohio State University Raj Jain 89 .References (Cont) q Cisco .cisco.

Thank You! The Ohio State University Raj Jain 90 .