Internet Routing. Shortest-Path Routing. Intra-Domain Routing and Traffic Engineering. Two-Tiered Internet Routing System

Transcription

1 Intra-Domain Routing and Traffic Engineering Review of Internet routing paradigm and routing algorithms/protocols Intra-domain routing design topology design, convergence, stability, Traffic Engineering (TE) MPLS and traffic engineering traffic engineering as networ-wide optimization problem TE through lin weight assignments Internet Routing So far we have focused on data plane operations Forwarding: data plane Directing a data pacet to an outgoing lin Individual router using a forwarding table Routing: control plane Computing the paths the pacets will follow Routers taling amongst themselves Individual router creating a forwarding table Two-Tiered Internet Routing System Shortest-Path Routing Path-selection model Interdomain routing: between ASes Routing policies based on business relationships No common metrics, and limited cooperation Destination-based Minimum hop count or sum of lin weights Dynamic vs. static lin weights BGP: policy-based, path-vector routing protocol Intradomain routing: within an AS Shortest-path routing based on lin metrics Routers all managed by a single institution OSPF and IS-IS: lin-state routing protocol RIP and EIGRP: distance-vector routing protocol 5

2 Distance Vector Routing: Bellman-Ford Define distances at each node x d x (y) = cost of least-cost path from x to y Update distances based on neighbors d x (y) = min {c(x,v) + d v (y)} over all neighbors v u v w E.g., RIP x 5 s y t z d u (z) = min{c(u,v) + d v (z), c(u,w) + d w (z)} Lin-State Routing: Dijstra s Algorithm Each router eeps trac of its incident lins Lin cost, and whether the lin is up or down Each router broadcasts the lin state To give every router a complete view of the graph Each router runs Dijstra s algorithm To compute shortest paths and forwarding table 5 E.g., OSPF 5 6 OSPF (Open Shortest Path First) Lin State Protocol lin costs between 0 and 65,55 Cisco recommendation - lin cost = /(lin capacity) rapid, loop-free convergence, scales well topology map at each node, route computation using Dijstra s algorithm OSPF advertisement carries one entry per neighbor router, advertisements flooded to entire Autonomous System multiple equal-cost paths allowed: flow equally split on all outgoing lins belonging to shortest paths IS IS (intermediate system-intermediate system) similar Dissemination Algorithm Converge Routing Protocols Lin State Flood lin state advertisements to all routers Dijstra s shortest path Fast due to flooding Distance Vector Update distances from neighbors distances Bellman-Ford shortest path Slow, due to count-to-infinity Protocols OSPF, IS-IS RIP, EIGRP BGP Path Vector Update paths based on neighbors paths Local policy to ran paths Slow, due to path exploration 7 8

3 Intra-Domain Routing Today Lin-state routing with static lin weights Static weights: avoid stability problems Lin state: faster reaction to topology changes Most common protocols in bacbones OSPF: Open Shortest Path First IS-IS: Intermediate System Intermediate System Some use of distance vector in enterprises RIP: Routing Information Protocol EIGRP: Enhanced Interior Gateway Routing Protocol Growing use of Multi-Protocol Label Switching What do Operators Worry About? Topology design Small propagation delay and low congestion Ability to tolerate node and lin failures Convergence delay Limiting the disruptions during topology changes E.g., by trying to achieve faster convergence Scalable routing designs Avoiding excessive protocol overhead E.g., by introducing hierarchy in routing Traffic engineering (focus today) Limiting propagation delay and congestion E.g., by carefully tuning the static lin weights 9 0 Topology Design: Intra-AS Topology Node: router Edge: lin Hub-and-spoe Bacbone ISP Topology Design: Points-of- Presence (PoPs) Inter-PoP lins Long distances High bandwidth Intra-PoP lins Short cables between racs or floors Aggregated bandwidth Lins to other networs Wide range of media and bandwidth Inter-PoP Intra-PoP Other networs

4 Topology Design: Abilene Internet Bacbone Convergence: Detecting Topology Changes Beaconing Periodic hello messages in both directions Detect a failure after a few missed hellos hello Performance trade-offs Detection speed Overhead on lin bandwidth and CPU Lielihood of false detection Convergence: Transient Disruptions Inconsistent lin-state database Some routers now about failure before others The shortest paths are no longer consistent Can cause transient forwarding loops 5 Convergence: Delay for Converging Sources of convergence delay Detection latency Flooding of lin-state information Shortest-path computation Creating the forwarding table Performance during convergence period Lost pacets due to blacholes and TTL expiry Looping pacets consuming resources Out-of-order pacets reaching the destination Very bad for VoIP, online gaming, and video 5 6

5 Convergence: Reducing Convergence Delay Faster detection Smaller hello timers Lin-layer technologies that can detect failures Faster flooding Flooding immediately Sending lin-state pacets with high-priority Faster computation Faster processors on the routers Incremental Dijstra algorithm Faster forwarding-table update Data structures supporting incremental updates Scalability: Overhead of Lin- State Protocols Protocol overhead depends on the topology Bandwidth: flooding of lin state advertisements Memory: storing the lin-state database Processing: computing the shortest paths Scalability: Improving the Scaling Properties Scalability: Introducing Hierarchy Through Areas Divide networ into regions Dijstra s shortest-path algorithm Simplest version: O(N ), where N is # of nodes Better algorithms: O(L*log(N)), where L is # lins Incremental algorithms: great for small changes Timers to pace operations Minimum time between LSAs for the same lin Minimum time between path computations More resources on the routers Routers with more CPU and memory area border router Bacbone (area 0) and non-bacbone areas Each area has its own lin-state database Advertise only path distances at area boundaries Area Area 0 Area Area Area 9 0

6 Limitations of Conventional Intra- Domain Routing Overhead of hop-by-hop forwarding Large routing tables and expensive loo-ups Paths depend only on the destination Rather than differentiating by source or class Only the shortest path(s) are used Even if a longer path has enough resources Transient disruptions during convergence Cannot easily prepare in advance for changes Limited control over paths after failure Depends on the lin weights and remaining graph Multi-Protocol Label Switching (MPLS) Motivating applications Small routing tables and fast loo-ups Virtual Private Networs Traffic engineering Path protection and fast reroute Key ideas of MPLS Label-switched path spans group of routers Explicit path set-up, including bacup paths Flexible mapping of data traffic to paths IP pacet Label Switching MPLS header MPLS header includes a label Label switching between MPLS-capable routers Regular IP Forwarding MPLS Label Distribution 7. IP 7... D e s t O u t D e s t O u t IP 7... IP 7... D e s t O u t IP Intf Label Dest Intf Label In In Out Out Request: 7. Intf Dest Intf Label In Out Out Mapping: Intf Label Dest Intf In In Out IP destination BISS 00: address FAN unchanged in pacet header!

7 Label Switched Path (LSP) Intf Dest Intf Label In Out Out IP 7... Intf Label Dest Intf Label In In Out Out Intf Label Dest Intf In In Out IP MPLS Forwarding: Example IP pacet destined to...5/ arrives to SF San Francisco has route for./6 next hop is LSP to New Yor IP...5 San Francisco 965 Santa Fe 06 0./6 New Yor 5 6 MPLS Forwarding Example San Francisco pre-pends MPLS header onto IP pacet, sends pacet to first transit router on path MPLS Forwarding Example because pacet arrived to Santa Fe with MPLS header, Santa Fe forwards it using MPLS forwarding table./6 New Yor./6 New Yor San Francisco Santa Fe San Francisco Santa Fe 7 8

8 MPLS Forwarding Example pacet arrives from penultimate router with label 0 egress router sees label 0, strips MPLS header egress router performs standard IP forwarding San Francisco Santa Fe IP./6 New Yor 9 MPLS: Pushing, Swapping, and Popping Pushing: add the initial in label Swapping: map in label to out label Popping: remove the out label IP edge A B Pushing IP IP R Swapping R R MPLS core IP Popping R IP C D 0 MPLS Header MPLS Header IP pacet encapsulated in MPLS header and sent down LSP Label EXP S TTL -bit MPLS Header IP Pacet IP pacet restored at end of LSP by egress router TTL adjusted by default label used to match pacet to LSP experimental bits carries pacet queuing priority (CoS) stacing bit: can build stacs of labels qoal: nested tunnels! time to live copied from IP TTL

9 MPLS Terminology LDP: Label Distribution Protocol LSP: Label Switched Path FEC: Forwarding Equivalence Class LSR: Label Switching Router LER: Label Edge Router (Useful term not in standards) MPLS multi-protocol both in terms of protocols it supports ABOVE and BELOW in protocol stac! Status of MPLS Deployed in practice Small control and data plane overhead in core Virtual Private Networs Traffic engineering and fast reroute Challenges Protocol complexity Configuration complexity Difficulty of collecting measurement data Continuing evolution Standards Operational practices and tools Traffic Engineering Goal: configure routes to meet traffic demands balanced load, low latency, service agreements operates at coarse timescales Not to adapt to short-term sudden traffic changes May tae potential failures into consideration Input to traffic engineering: Topology: connectivity & capacity of routers & lins Traffic matrix: offered load between points in the networ Traffic Engineering: networ-wide optimization Subject to protocol mechanisms, configurable parameters and other practical constraints,. Traffic Engineering under Shortest Path Routing: Tuning Lin Weights Problem: congestion along the blue path Second or third lin on the path is overloaded Solution: move some traffic to bottom path E.g., by decreasing the weight of the second lin 5 5 6

10 Traffic Engineering Framewor Traffic Engineering Framewor (cont d) Basic Requirements Knowledge of Topology Connectivity and capacities of routers/lins of a networ Traffic Matrix (average) traffic demand between difference ingress/egress points of a networ Instrumentation Topology: monitoring of the routing protocols Traffic matrix: fine-grained traffic measurement and inference, for example, via SNMP edge measurements + routing tables networ tomography pacet sampling Traffic Engineering as Networ-Wide Optimization Networ-wide models Networ topology: graph (V,E), c ij : capacity of lin (i,j) Traffic Matrix: K set of (ingress/egress) source-destination pair demands K, d demand, s source, t destination Optimization criteria, e.g., minimize maximum utilization, minimize sum of lin utilization eep utilizations below 60% 7 8 Traffic Engineering as a Global Optimization Problem topology G = (V,E) c ij capacity of lin (i,j) E K set of origin-destination flows (demands) K, d demand, s source, t destination F ij: traffic load of O-D flow routed on lin (i,j) F ij := F ij: total load (of all demands) on lin (i,j) Optimization objective function: Φ({f ij, c ij }) e.g., Φ({f ij, c ij }):= max {(i,j) E} {f ij /c ij } or Φ({f ij, c ij }):= (i,j) E f ij /c ij Traffic Engineering as a Global Optimization Problem (cont d) Objective function: minimize Φ({f ij, c ij }) Constraints: -- flow conservation: total outflow vs. total inflow) j :( i, j ) E F 0, i s, i t, K = d, i = s, i t K d, i s, i = t, K ij F, j :( j, i ) E ji -- capacity and (non-negative load) constraints K Fij F ij 0 c ij, ( i, j) E 9 0

11 Traffic Engineering Example: minimize maximum lin utilization Minimize Φ({f ij, c ij }):= max {(i,j) E} {f ij /c ij } Multi-commodity flow problem There exists polynomial time solutions to the problem Equivalent linear programming formulation α maximum lin utilization Let X ij:=f ij/d X ij denotes fraction of demand on lin (i,j), X ij [0,] s. t. min α 0 X j:( i, j) E K d ij X Traffic Engineering: LP Formulation X ij ij c α, ij j:( j, i) E X ji 0, i s, t, K =, i s K, i t K ( i, j) E Traffic Engineering w/ MPLS We can set up label-switched paths (LSPs) between origin-destination pairs to realize the optimal TE traffic load distributions Let {X * ij} be the optimal solutions If for a given (corresponding to a given O-D pair), X * ij = 0 or, then we set up one LSP (or tunnels) for the O-D pair Otherwise, traffic load for flow (demand) is carried over multiple paths, we need to set up multiple LSPs (or tunnels) for the given O-D pair In general, traffic split among multiple LSPs are not equal! worst-case complexity: O(N^E) LSPs/tunnels needed Traffic Engineering w/o MPLS Can we perform traffic engineering without MPLS? we need to use shortest path routing But shortest paths are defined based on lin weights TE becomes lin weight assignment problem! Other constraints we need to tae into account destination-based routing: not <src, dst> pair-based! multiple shortest paths ( equal-cost paths, ECPs) may exist and can be used for load-balancing

12 Shortest Path Routing and Lin Weight Assignment Problem Key Problem: how to assign lin weights to optimize TE objectives under conventional lin-state (shortest path) routing paradigm? Key Insight: traffic engineering optimization is closely related to optimal lin weight assignment using shortest path routing (with some caveats!) The relationship comes from duality properties of linear programming optimal lin weight assignment problem is a dual problem to the optimal traffic engineering problem!

13 Duality of Linear Programming Primal minimize subject to Ax x c T x = b 0, Dual maximize subject to y T y T b A c y s are Lagrange multipliers for equality constraints Ax=b; z 0 Lagrange multipliers for inequality constraints x 0 Lagrange function L(x,y,z) := c T x -y T (Ax-b)-z T x c T x, if x feasible Lagrange dual: g(y,z): = inf x 0 L(x,y,z) c T x * (x * optimal solution) g(y,z) = y T b, if c T -y T A z T 0, or y T A c T as z 0 T Complementary Slacness. Let x and y be feasible solutions. A necessary and sufficient condition for them to be optimal is that for all i. x i > 0 y T A i = c i. x i = 0 y T A i < c i Here A i is i-th column of A

14

15

16 Lin Weight Assignment wors for rich set of cost functions example: Φ = Φ d X ( i, j ) E ij ( ij ) K where Φ ij are piecewise linear 6 6