Routing architecture and forwarding

Transcription

1 DD2490 p Routing architecture and forwarding & Intro to Homework 4 Olof Hagsand KTH /CSC 1

2 Connecting devices Connecting devices Networking devices Internetworking devices Hub/ Hub/ Repeater Bridge/ Switch Router Application gateway L1 L2 L3 L4-L7 2

3 IEEE 802 vs IPv4 addresses vendor code vendor assigned IEEE Group/ Individual bit Global/ Localbit 00:0E:35:64:E9:E7 netid hostid IPv4 addr IP addresses have an hierarchical structure: the netid is used to subnet. MAC addresses have a flat structure.

4 Routing vs bridging Bridging - forwarding on layer 2 A MAC address/id has a flat structure Many nodes -> large forwarding tables Broadcast reaches all nodes Simple to configure and manage, cheaper Loops detected by spanning tree protocol Routing forwarding on layer 3 The netid of the IP addresses can be aggregated Many nodes -> smaller forwarding tables than bridging Routers partition broadcast domains Routing is more difficult to configure Loops detected by routing protocols and TTL decrementation 4

5 What does a router do? Packet forwarding Not only IPv4: IPv6, MPLS, Bridging/VLAN, Tunneling,... Filter packets - Access lists Classification Metering/Policing/Shaping Compute routes: build forwarding table In the background: routing In real-time: forwarding Classifier Lookup Metering Shaping 5 ACL - Access Lists for dropping packets Metering - Measure traffic characteristics of a 'flow': rate in terms of packet-per-second and bits-per-second Policing - If the rate is higher than a threshold, drop packets. Alternatively, packets can be 'marked' as out-of-norm and potentially dropped at a later stage. For example, packets being marked as out-of-profile atthe edge may be dropped within a network if congestion is detected. Shaping - Actively changing the rate of traffic, making it comply to a specific rate. Often use token-bucket / leaky-bucket mechanisms or most simply a queue. Shaping is more complex to do than policing.

6 Inside a software-based router CPU RIB Buffer Memory Shared bus backplane Line Card Line Card Line Card This is a regular computer architecture (eg PC) Every packet goes twice over the shared bus Constrained by Bus and memory bandwidth (per byte cost) And CPU cycles (per packet cost) 6 First generation routers were in fact regular computers

7 Inside a hardware-based router Switched backplane Line Card Buffer Memory Line Card Buffer Memory CPU Card CPU Line Card Buffer Memory Line Card Buffer Memory forwarder forwarder forwarder forwarder RIB Multiple simultaneous transfers over the backplane Specialized hardware: ASICs (Application Specific IC) Wirespeed at 100 Gb/s and beyond 7

8 Fast path, slow path Control Processor CPU Memory Routing Table Slow path Line Card Line Card Fast path Fast path If line cards can determine outgoing port Slow path Line Card Control processor must determine outgoing port Line Card 8

9 Routing table lookup Longest prefix first Divide table in 32 buckets - one for each netmask length Match destination with longest prefixes first SW algorithms: tree, binary trees, tries (different data structures) HW support: TCAMs Content Addressable Memory Masklen 0 1 Netid Netid destination IP address 9 The shift to classless addressing led to changes to the routing table organization and routing algorithms since the netid length is variable. You may have several matches and you need to select the longest prefix. How much time can you spend on lookups? Example: 10 Gb/s line rate and 40 bytes packet size: ns to do sequential lookup (comparable with SRAM memory speed) The example above shows a 'bucket' approach to lookup. Organize the routing table in buckets based on prefix length and check in longer buckets first. This is only slightly better than a linear search.

10 Using a Trie for lookup Binary tree Nodes are prefixes Left branch represents 0 in the string Right branch represents 1 * a 0* 1* a * b 10* c 01* d 110* e 0010 f 0110 g * 01* c 10* b 11* 000* 011* 110* d 0010 e 0110 f g A much better algorithm is to organize the routing table into a binary trie. There are at most four lookups required in the example above.

11 Elimination of Internal Prefixes No overlapping prefixes Prefix expansion with leaf pushing Simplifies lookup at expense of larger memory * a * b 10* c 01* d 110* e 0010 f 0110 g * 01* 10* 11* a a e a c c f g b b b b d d a a 11 Tries are often modified for optimal lookup using compression and expansion. The above is one example. Only one lookup required.

12 Linear Search on Values TCAM Ternary Content-Addressable Memory Fully associative memory Three values for each bit 0, 1, and x (don t care) Compare input with all words in parallel First match gives the result Up to 100 million searches per second input = = = = = = = TCAM 0010 g 0110 f 0111 e 110x d 01xx c 10xx b xxxx a a * b 10* c 01* d 110* e 0010 f 0110 g Example of hardware supported lookups - not a software-based method.

13 Packet classification Map a packet to a class Class defined by filters, usually a 5-tuple: <source IP, destination IP, source port, destination port, protocol> For example, all packets: From subnet N To TCP port 80 on web-server S From subnet N to port 666 on subnet M Applications: Firewall & NAT Blocking Accounting Policy routing QoS metering, policing, DiffServ marking, This is called Forwarding Equivalence Class (FEC) in MPLS-speak.

14 Examples of commercial routers Capacity: 80Gb/s Power: 2.6kW Capacity: 1.28 Tb/s Power: 4.7 kw 6ft 3ft 2ft Cisco ft Juniper M This slide is somewhat dated,...

15 Here is one of the SUNET core-routers in Stockholm, a Juniper T- router. 15

16 Juniper J-series J4300 Routers used in labs Emulates M/T series Full routing software 16

17 Open source routing Linux and BSD platforms Most routing protocols exist as open source projects (eg Quagga) But PC hardware has traditionally been a limiting factor But now up to 4x12 core CPUs, inter-processor buses (HT, QPI), non-uniform memory (numa), I/O buses (PCI-E), 10Gbps NICs enables 10s of gigabit forwarding speeds. Example: the Bifrost open source router (UU/KTH) 17 A router as the above routes all Uppsala University IPv4 and IPv6 traffic at 2x speed.

18 Example: PC routing architecture DDR3 DDR3 DDR3 CPU 0 CPU 1 CPU 2 CPU 3 QPI CPU 4 CPU 5 CPU 6 CPU 7 DDR3 DDR3 DDR3 PCI-E x16 x4 x16 QPI I/O Handler (North Bridge) QPI I/O Handler (North Bridge) PCI-E x16 x4 x16 Multi-core CPUs: (Intel Nehalem) 8 cores, 16 with 'Hyperthreading' Multi-channel: each network card has 8 DMA queues NUMA: Non-local memory (many memory banks) Inter-processor bus: QPI 2.4GHz ~76 GB/s Memory: 1066 DDR3 68 GB/s x3 channels I/O Bus: PCI-E gen2 x1 ~4GB/s: x16 ~64GB/s 18 New Intel (sandy bridge) and AMD (bulldozer) provides even more CPU cores.

19 Homework 4 Make a programming assignment in C Part 1: Print out IPv4 destination address Part 2: Make an IPv4 forwarding lookup Mandatory: 3 bonus points Part 3: Same as part-2 with non-trivial lookup (time limit) Optional: 2 bonus points 19

20 Homework 4: Part 1 You should read an Ethernet frame, identify it as an IPv4 packet, and print the IPv4 destination address. Input: Ethernet packet. Example: c d ff01 8ffc 0a a e802 c04b e89 339a 0786 d0ff 0009 Output: IPv4 address. Example: Errors: Error: packet too short: length of frame in bytes Error: Not ipv4 payload: payload type 20

21 Homework 4: Part 2 The program should read a forwarding table and an Ethernet packet and extract the destination IP address, make a lookup in the forwarding table, and write the outgoing interface name. The assignment is a step towards a full forwarding but lacks several sanity checks, MAC address lookups and ARP. It is intended to illustrate how to inspect packet header, the use of pointers, buffers, and IP longest prefix match. The program should do the following: Read a routing table from stdin. The routing table consists of a list of prefix, nexthop interface triples. Read a single Ethernet (RFC894) packet from stdin. Verify that the packet is long enough to contain an EThernet and IPv4 header Verify that the Ethernet payload type is 0x0800 (IPv4) Verify that the IP version field is 4 Extract the destination address from the IPv4 header and make a longest prefix match lookup and return the outgoing interface name. Example: Input fib /24 e1 fib /24 e2 fib /24 e3 fib /0 e1 Example Output: e2 input d ff01 8ffc 0a a e802 c04b e89 339a 0786 d0ff

22 Homework 4: Part 3 Same as part 1 with two changes: 1) Read not only one but several packets 2) Time-limit on lookup The time-limit is set so that you cannot just have a simple linear lookup. The test-case used Part 3 is optional 22

23 Homework 4: Kattis If you have registered, you will get a Kattis account Use the link on the homework page and login Submit by selecting language: C Select problem: forwarding (part 1), forwarding2 (part2), forwarding3 (part3) upload the file Submit You can see the status on the web-page Compile-error Runtime error Wrong output OK You will also get a mail Submit solution electronically, or on paper lab assistants or course leader before the deadline. Append a receipt that you passed both forwarding and forwarding2 (& forwarding 3) test of Kattis. 23 Read the instructions for the kattis assignment thoroughly. Kattis is a machine and is very picky about details: Extra spaces, capital letters, etc, are significant.

24 Extracting correct info The Ethernet header is 14 bytes payload type is in bytes IPv4 is 0x0800 The IP header is 20 bytes (without options) The destination IP address is in bytes struct ethhdr{ }; char uint16_t pt; struct iphdr{ }; da[6], sa[6]; unsigned int ip_v:4,ip_hl:4; /* version, header length uint8_t ip_tos; /* type of service */ uint16_t ip_len; /* total length */ uint16_t ip_id; /* identification */ uint16_t ip_off; /* fragment offset field */ uint8_t ip_ttl; /* time to live */ uint8_t ip_p; /* protocol */ uint16_t ip_sum; /* checksum */ uint32_t ip_src, ip_dst; /* source and dest address */ 24

25 Byte ordering / Endianness CPU:s represent numbers they load/store from memory differently Most significant byte in first byte: Big-endian (Big end first) Most significant byte in last byte: Little-endian There is also middle-endian and bi-endian Intel PCs are little-endian. The Kattis machine is big-endian (Sun Sparc) Register 0A0B0C0D Memory Register 0A0B0C0D Little-Endian 0A 0B 0C 0D n n+1 n+2 n+3 0A 0B 0C 0D Big-Endian 25

26 Network byte order The way the CPU stores/loads numbers from memory is called host byte order But in communication system, we sometimes have to transfer numbers in binary format (character arrays is not a problem) We have to agree on a format to encode numbers This is called network-byte order In IP network-byte order is big-endian Therefore, in portable code, if you transfer binary numbers between nodes, always translate between host-byte order and network-byte order. BSD has the following help functions: htonl, ntohl (4-byte numbers) htons, ntohs (2-byte numbers) 26

27 Alignment Data structures must be aligned in memory when accessed as several bytes. In particular 2-byte, 4-byte, 8-byte numbers must be aligned on word boundaries Otherwise a bus error occurs (in serious cases, eg SPARC) Or a performance degradation (as in x86) Typically, 2-byte numbers must be 2-byte aligned 4-byte numbers must be 4-byte aligned Etc In Eth+IP, the Eth header is 14 bytes which makes the IP header misaligned (actually, the fields of the IP header) 27

28 Alignment example Memory A 0B 0C 0D 0A0B0C0D OK Memory A 0B 0C 0D 0A0B0C0D BUS ERROR! 28

29 More info See C-programming hints on course web Ask course instructor Or Lab assistants during labs There are many C-programming tutorials on the web 29