Lecture on Computer Networks

Transcription

1 Lecture on Computer Networks Historical Development Copyright (c) 28 Dr. Thomas Haenselmann (Saarland University, Germany). Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. time

2 Local Area Networks Overview 5. Packet Switching 5.2 Virtual Circuits vs. Datagrams 5.3 Routing in Unicast Networks 5.4 Routing in Multicast Networks 5.5 Congestion Control 5.6 Examples: IP Version 4, IP Version 6, ATM

3 The Network Layer in the OSI reference model Application Presentation Session Transport Network Data Link Physical router physical medium A physical medium B

4 ISO Definition for the Network Layer The network layer provides the ability to establish, operate and terminate network connections between open systems over intermediate systems. The network layer provides independence from routing and switching decisions.

5 Functions of the Network Layer Routing and switching of packets Multiplexing of end-to-end connections over a layer-2 connection Packet segmentation ( fragmentation") In connection-oriented communication the network layer additionally provides connection establishment and termination error detection and error correction (end-to-end) guaranteeing the order of the packets flow control (end-to-end) Heterogeneous subnetworks can be interconnected by a network layer instance ("Internetworking").

6 Overview of routing- and addressing schemes (from Wikipedia) Unicast: The most common concept of an IP address is a unicast address. It normally refers to a single sender or a single receiver, and can be used for both sending and receiving. Usually, a unicast address is associated with a single device or host, but it is not a one-to-one correspondence. Some individual PCs have several distinct unicast addresses, each for its own distinct purpose. Sending the same data to multiple unicast addresses requires the sender to send all the data many times over, once for each recipient.

7 Overview of routing- and addressing schemes (from Wikipedia) Broadcast: Sending data to all possible destinations (an "all-hosts broadcast") permits the sender to send the data only once, and all receivers can copy it. In the IP protocol, represents a limited local broadcast. In addition, a directed (limited) broadcast can be made by combining the network prefix with a host suffix composed entirely of binary s. For example, to send to all addresses within a network with the prefix 92..2, the directed broadcast IP address is (assuming the netmask is ).

8 Overview of routing- and addressing schemes Multicast: A multicast address is associated with a group of interested receivers. According to RFC 37, addresses to are designated as multicast addresses. This range was formerly called "Class D." The sender sends a single datagram (from the sender's unicast address) to the multicast address, and the intermediary routers take care of making copies and sending them to all receivers that have registered their interest in data from that sender.

9 Overview of routing- and addressing schemes Anycast: Like broadcast and multicast, anycast is a one-to-many routing topology. However, the data stream is not transmitted to all receivers, just the one which the router decides is the "closest" in the network. Anycast is useful for global load balancing, operating via the shortest-path metric of BGP routing. It is most commonly used in DNS, but does not take into account congestion or other nonpath metrics.

10 Virtual Circuits vs. Datagrams Virtual Circuit The path through the network is determined when the connection is established, i.e., for each new virtual connection a routing decision takes place in each node only once. The entire traffic flowing over this virtual connection takes the same path through the network. Datagram Every packet contains the full address of the destination host. When a packet arrives at an intermediate node the destination address is used to determine the outgoing link for the next hop.

11 Virtual Circuit "Perfect" channel through the network Error control (bit errors, lost and duplicated packets) Flow control Order of the messages maintained Phases Connection establishment Data transmission Connection termination Advantages Low overhead for addresses in the data transmission phase Low computational overhead for routing in the data phase High quality of the arriving packet stream

12 Implementation of Virtual Circuits Tables with status information on all existing virtual circuits are maintained in each node. (a) Example of a subnetwork: H H H o s t H B C H A D E F (b) Eight virtual connections through this subnetwork: H Starting from A Starting from B ABCD AEFD 2 ABFD 3 AEC 4 AECDFB BCD BAE 2 BF H

13 Status Information in the Nodes (c) Routing tables in the nodes Wide Area Networks and Routing H B H H H H B E E B E E 2 3 Input Output A B Incoming IMP or Host Incoming virtual connection 2 C C A F F H A H H A H F C 2 D D H D B B E E F D D H B E B B D E 2 3 F H C C A A A A D H H H H F C C F F C

14 The Datagram Every packet (datagram) is considered an isolated unit (like a telegram): Full destination address contained in every packet Packets can arrive out of order No error control, no flow control in layer 3 Advantages Simpler than virtual circuits, therefore easier to implement No connection establishment and termination phases: low overhead for short-lived connections More reliable since there are no status clean-up and recovery problems when a node or link fails Better suitable for internetworking of heterogeneous subnetworks

15 Routing in Unicast Networks Topology We assume that the network topology is a graph (note that in LANs with a broadcast topology (bus, ring) routing is not required!).

16 Routing Algorithms () Task To route the packets through the network from the source end system to the destination end system. The routing algorithm decides to which outgoing link of each router an in-coming packet is transferred. Desirable characteristics of a routing algorithm Correct Simple Robust in case of node or link loss Fair Optimal (finds the best route, causes minimal overhead)

17 Routing Algorithms (2) These design criteria are in conflict. In practice a good approximation is: minimization of hops from the source to the destination.

18 Classification of Routing Algorithms. Static (non-adaptive) Routing No consideration of the current network conditions For all i, j the routes between i and j are determined before the start-up of the network. No change during operation 2. Adaptive Routing Decisions are based on the current network topology (and sometimes load) Continuous measurement of the topology and the traffic Can be further classified into - centralized routing algorithms - isolated routing algorithms - distributed routing algorithms

19 Static Routing In static routing the entire topology of the network is known to a central node. It computes the optimal paths for each pair (i,j) of nodes, generates routing tables for the individual nodes and sends them out. Static routing is only possible if the network is relatively small and changes to the topology are rare. Multipath Routing Alternative routes are computed for each pair of nodes (i,j). The probability of use depends on the relative weights. Multipath routing is more reliable in case of node or link failures and balances the load better. However it is more complex than single-path routing.

20 Multipath Routing () Every node contains a routing table with one row for every destination: D O W O 2 W 2 O n W n D O i destination i th best outgoing link W i weight for O i W i is the probability that O i is used. n W i i= =

21 Multipath Routing (2) Selection of an alternative: generate a random number z ( <= z <= ) choose O if <= z < W choose O 2 if W <= z < W + W 2... choose O n if (W + W W n- ) <= z <=

22 Static Routing: Example Topology Topology of the example network A B C D E F G H I J K L We consider the paths beginning at node J.

23 Static Routing: Example Table Static routing table with alternative paths for node j Des. st choice 2 nd choice 3 rd choice A A.63 I.2 H.6 B A.46 H.3 I.23 C A.34 I.33 H.33 D H.5 A.25 I.25 E A.4 I.4 H.2 F A.34 H.33 I.33 G H.46 A.3 K.23 H H.63 K.2 A.6 I I.65 A.22 H.3 - K K.67 H.22 A. L K.42 H.42 A.6

24 Determination of Routing Tables For static routing the routing tables are precomputed by the network operator. Before network start-up they are loaded into the nodes and not changed anymore. Characteristics simple good results for a constant topology and constant network traffic But: inappropriate for strongly varying traffic and changes in topology inappropriate for large networks (does not scale well) Still occasionally used in practice. The network operator always knows the entire topology. He/she can use Dijkstra s shortest path" algorithm once for each node for the construction of the routing tables.

25 Centralized Adaptive Routing () Principle There is a Routing Control Center (RCC) in the network. Each node periodically sends status information to the RCC, for example - the list of immediate neighbors - current queue lengths - the current utilization of its links. The RCC collects the information and computes the optimal path for each pair of nodes, computes the individual routing tables and distributes them to the nodes.

26 Centralized Adaptive Routing (2) Characteristics The RCC has complete information => decisions are optimal. The individual nodes don t have to do the routing computation. But: Route computation has to take place frequently (say, once per minute). There will be a traffic concentration in the proximity of the RCC, thus a performance bottleneck. The technology is not robust: the RCC is a single point of failure. The algorithm fails when the network gets partitioned. The individual nodes receive new routing tables at different times => inconsistencies and thus "routing loops" will occur.

27 Centralized Adaptive Routing: (side effect of asym. metrics) recv from x, y, z w recv from x, y, z w send z +e x send send z +e 2+e x send e y y send e send e send recv from x, y, z w z y send e 2+e +e x send send recv from x, y, z w z +e 2+e y send e x send

28 Centralized Adaptive Routing: Algorithm (performed by node u) D(v)=cost(u,v) INIT N'={u} for all nodes v if v is a neighbor of u then D(v)=cost(u,v) else D(v)=infinity LOOP find w not in N' such that D(w) is a minimum add w to N' update D(v) for each neighbor v of w not in N': D(v)=min(D(v), D(w)+c(w,v)) UNTIL N'=N

29 Isolated Adaptive Routing Principle No exchange of routing information between nodes Decisions are based on local information only Examples of algorithms Backward Learning Flooding

30 Algorithm "Backward Learning" A node "learns" from the arriving packets: packet (..., S, H,... ) with S = source node, H = hop counter, the packet is received on link L => S is reachable over L in H hops. Routing table in the node: Each entry is a triple (destination node, outgoing link, H min ) Updating of the routing table: Node receives packet (..., S, H...) on link L if not(s in table) then add(s,l,h) else if H < H min then update(s,l,h)

31 Backward Learning: Example P L S L 2 D P2 P(..., S,4,...) -> add(s,l,4) P2(..., S,3,...) -> update(s,l2,3)

32 Backward Learning: Path Degradation Problem Algorithm does not adapt to path degradations. Solution Periodic deletion of routing tables. A new learning period begins. But: how often? too often: network is in the learning phase most of the time too seldom: slow reaction to degradations Also: We cannot send data to a node from which we have never seen a packet. Approach used by routers and LAN bridges (empty entry => flood packet)

33 Algorithm "Flooding" An incoming packet is forwarded on all outgoing links except the one it came from. A B C

34 Flooding: Stopping the Packet Flow Problem: explosion of the number of packet copies in the network Mitigating the problem add a hop counter to the packet header initialize it with the diameter of the network (= longest path in the network (worst case)) decrement the hop counter on each hop counter = : packet is dropped by the router Characteristics of Flooding very robust, very simple, but large number of copies, heavy network load => employed only for special applications or in a first phase of other routing algorithms to establish initial status information

35 Flooding example in sensor networks: Directed Diffusion Aim: Get measurements from the nodes of a sensor network Therefore, the sink node sends out a so-called interest by flooding Every node stores an interest and forwards it to its neighbors As the diameter of the network is unknown: keep track which neighbor got or sent a copy of the interest and don't send twice struct interest type = walking moose // taken from an app. specific set frequency = s // time between 2 messages location = rect[(6,4),(8,8)] // only within this area timestamp = 4:25:35 // birth of this interest expiresat = 4:3: // no answers after this time

36 Flooding example in sensor networks: Directed Diffusion Answering an interest A node N that can satisfy the interest by its observations answers by flooding as well. Consequence: Huge network load, so answer in large intervals only. One answer reaches the sink on the shortest link (why?). It then reinforces the interest via the node reporting first. This node does so as well. So the reinforcement gets to N on the optimal path P. From now on, N will answer via P until timeout. + approach for routing by content, not by identities - lots of status information, feasible for dedicated applications only.

37 Distributed Routing () Principle The nodes explicitly exchange routing information with their neighbors: Each node knows its distance to each neighbor: - number of hops (= ) - delay (round-trip time) - queue length, etc. Each node periodically sends a list with his estimated distances to all known destination nodes to his neighbors. node X receives a list E from neighbor Y distance (X, Y) = e distance (Y, Z) = E(Z) => distance (X, Z) over Y is E(Z) + e The table with these distances is called distance vector. The algorithm is thus called distance vector routing.

38 Distributed Routing (2) Example A B C D E F G H I J K L We consider the distances known to node J.

39 Distributed Routing (3) A B C D E F G H I J K L A JA delay=8 I JI delay= H JH delay=2 delay=6 Right column: newly determined distances at node J after receiving the distance vectors from the neighbors K JK new DV at J 8 A 2 A 28 I 2 H 7 I 3 I 8 H 2 H I - 6 K 5 K

40 Hierarchical Routing With no hierarchie, the size of the routing tables is proportional to the size of the network: large memory requirement in the nodes considerable CPU time for searching the tables much bandwidth for the exchange of routing information Hierarchical routing helps to solve these problems: Nodes are grouped into regions Each node knows - all details of his region - his routes to all other regions In the Internet a region is a subset of the IP address space. Disadvantage: globally optimal decisions are no longer possible.

41 Example for Hierarchical Routing B region region 2 2A 2B full table for node A A C 2C 2D 4B 5B 5C 3A 3B 5A 5D 4A 4C 5E region 3 region 4 region 5 hierarchical table for node A DES. LIN HOP A E- - B B C C 2 B 2 3 C 2 4 C 3 5 C 4 DES. A B C 2A 2B 2C 2D 3A 3B 4A 4B 4C 5A 5B 5C 5D 5E LINE - B C B B B B C C C C C C C B C C HOP

42 Routing in the Internet Distance Vector Routing In the early years of the Internet the most commonly used procedure was an adaptive distributed procedure on the basis of distance vectors (distance vector routing). The employed protocol is called RIP (Routing Information Protocol). With RIP all Internet routers periodically exchange distance vector messages and update their routing tables accordingly.

43 Routing in the Internet Algorithm (Distance Vector Routing) INIT (at node X) for all destinations y in N: Dx(y)=cost(x,y) // y not neighbor? cost=inf. for each neighbor w Dw(y) = infinity for all destinations y in N for each neighbor w send distance vector Dx=[Dx(y): y in N] to w LOOP FOREVER wait (until link cost changes or get new D) for each y in N: Dx(y)=min_v{c(x,v)+Dv(y)} if Dx(y) changed for any destination y send Dx=[Dx(y): y in N] to all neighbors

44 Example for Distance Vector Routing () A From A to B C D link ab ab ad ab cost 2 B From B to A C D link ab bc bc ad cost 2 bc (a) node E has just been added to the network C From C to A B D E D From D to A B C E cd link bc bc cd ce link ad cd cd de cost 2 cost 2 ce de E From E to C D link ce de cost

45 Example for Distance Vector Routing (2) B C A From A to B C D E link ab ab ad ad ab cost 2 2 From B to A C D E link cost From C to ab A bc bc B bc bc 2 2 D E ad (b) after an exchange of RIP messages D From D to A B C E cd link bc bc cd ce link ad cd cd de cost 2 cost 2 de ce E From E to A B C D link de ce ce de cost 2 2

46 Distance Vector Routing: Problems Good news travels fast: We learn quickly about better links Bad news travels slow: What happens with a deteriorated link? 4 y 6 y x 5 z x 5 z

47 OSPF Routing Another important and widely used routing algorithm on the Internet is OSPF (Open Shortest Path First). The basic idea is that all nodes know the entire network topology at all times and can thus compute all optimal paths locally. If the topology changes, the nodes exchange topology update messages. Each node maintains a local database of the entire topology, called the link state database. The optimal paths to all destinations can be computed locally with Dijkstra s Shortest Path algorithm (Shortest Path First = SPF). In the Internet slang the algorithm is thus called Open Shortest Path First. Algorithms of this class are called link state routing algorithms.

48 Example for OSPF Routing B C ab bc ce E A D cd de ad

49 Example for OSPF Routing Routing is done at layer 3, the Network Layer. At runtime a router forwards the incoming packets on the basis of a routing table. Hierachical routing helps to solve the scalablility problem for large networks. A routing algorithm computes the entries of the routing tables. Static routing can only be used for small networks with little change in topology and traffic. Adaptive routing with a centralized routing control center has the disadvantage of a central performance bottleneck and a central point of failure. The Internet uses adaptive distributed routing algorithms. The most popular ones are RIP (older, based on distance vectors) and OSPF (currently used, based on a link state database).