CSE 123b Graduate Networking

Transcription

1 CSE 123b Graduate Networking Spring 2004 Last Lecture: Potpourri Stefan Savage

2 Final Check Studentlink to be sure, but I believe that its Wed Jan 9 th 11:30-2:30 in this room. Closed book You can bring on 8.5x11 sheet of paper and you can do just about anything to that piece of paper you want (i.e. write on both sides, print on it, etc) You can t tape your textbook to the sheet of paper though Or your laptop Same style as midterm FYI: You may be asked questions about project Will cover entire class

3 Review session Still working on getting a room Most likely on Monday at 4-5pm I ll post details on the Webboard as soon as I know for sure

4 Stuff we can look at today Options Skim IPv6 Some multimedia + VOIP Example of Network measurement tricks How high-speed routers are designed

5 IPv6 Originally motivated by crisis in IP address space 32 bits not enough if everything gets an IP address Subnetting and CIDR helped alot, but not a long-term solution (mobile phones) Solution: Increase the size of IP addresses Originally to 64, then to 128 bits Requires changes to IP header If we re going to change the header, might as well change other aspects of IP Rule of thumb: IPv6 = IPv4 + big addresses + IPSEC, Mobile IP + DHCP + ARP + improvements to these

6 IPv6 (2) IP addresses are 128 bits Unicast, multicast, local, etc. 1234:ABCD:1234:ABCD:1234:ABCD:1234:ABCD Registry, provider, subscriber, subnetwork, interface Interface ID used as lowest 6 bytes Simplified headers 40 bytes (20 for IPv4) Use header extensions for options Autoconfiguration Multiple addresses per host (link-local address) FlowLabel, priority, security, mobility

7 IPv6 and IPv4 Interoperability Dual-stack operation IPv6 nodes that support both IPv4 and IPv6 Tunneling IPv6 packets encapsulated within IPv4 packets End-points speak IPv6, but use IPv4 packet to use standard Internet routing Easy mapping if IPv6 embeds IPv4 address, otherwise need to configure a table

8 Multimedia & IP Telephony Two sets of issues How much does the network help in making these applications possible? How does the application participate?

9 Network requirements Audio/video is a real-time service in the sense that data must be received by a deadline to be useful Microphone Sampler, A D converter Internet Buffer, D A Speaker Variable bandwidth and delay (jitter) Real-time apps need assurances from the network Q: What assurances does playback require? VOIP?

10 Network Support for Playback Bandwidth Delay Loss There must be enough on average But we can tolerate to short term fluctuations Ideally it would be fixed But we can tolerate some variation (jitter) Ideally there would be none But we can tolerate some losses

11 Example: Delay and Jitter 3 90% 97% 98% 99% Packets (%) Delay (milliseconds)

12 Tolerating Jitter with Buffering Sequence number Packet generation Network delay Buffer Packet arrival Playback Time Buffer before playout so that most late samples will have arrived

13 Application requirements Application variations Rigid fixed playback point (low jitter) Adaptive application can vary playback point (jitter ok) Tolerant can tolerate interrupt/degradation Intolerant can t In reality only two classes Rigid and intolerant applications (e.g. telemedicine, perhaps ip telephony) Adaptive and tolerant (e.g. RealPlayer)

14 How can network provide guarantees? Default Internet -> best effort Two new service models Integrated Services» Multiple service classes with explicit service specification per flow» I need 500kbps, with a maximum delay of 50ms no more variation than 5ms» Applications explicitly reserve capacity along path to destination (RSVP protocol)» Lots of state, but precise guarantees Differentiated services» Multiple service classes» I need my voice packets to be put in the platinum class of service» Packets are marked to indicate what class they should be in» More scalable, but not guaranteed service In all cases routers scheduling modified (no longer FIFO) Neither widely deployed for general use

15 Token Buckets Common, simple descriptor Fill rate R tokens/sec Use tokens to send bits Average bandwidth is R bps Maximum burst is B bits Bucket size B tokens Sending drains tokens

16 Using Token bucket traffic conditioning Operation Token bucket size: b (maximum number of tokens) Token bucket rate: r (rate at which new tokens are provided) Need k tokens to send a packet of k bytes Applications Shaping (delay packets until k tokens appear)» Long term rate is limited to r, short term bursts to b» Over some interval T, traffic is limited to b+r*t Policing (drop packets if insufficient tokens) Buffer management (mark packets and transmit if > k)» Drop marked (i.e. out-of-spec) packets during congestion

17 Token Bucket Operation Tokens Tokens Tokens Overflow Packet Enough tokens packet goes through, tokens removed Packet Not enough tokens wait for tokens to accumulate, drop or mark packet

18 Multimedia protocols Data transport RTP (Real-time protocol)» UDP-based protocol that carries encoded data, specifies the encoding, provide time synchronization and the ability for the receiver to order RTCP (Real-time control protocol)» For determining how well data is being received and adapting sending rate or coding (e.g. congestion control) Real (RDP) and Microsoft also have proprietary protocols that provide these functions Signaling RTSP (Real-time streaming protocol)» Looks just like HTTP, but is for informing server about multimedia file you want to view (where to get the audio, where to get the video, what bit rate, etc) SIP (Session Initiation protocol) and H.323» For setting up phone calls» SIP+RTP is becoming a dominant model (flexible allows new services)

19 SIP Clients and Servers SIP uses client/server architecture Elements: SIP User Agents (SIP Phones) SIP Servers (Proxy or Redirect - used to locate SIP users or to forward messages.)» Can be stateless or stateful SIP Gateways:» To normal phone system (PSTN) for telephony interworking» To H.323 for IP Telephony interworking Client - originates message Server - responds to or forwards message

20 SIP Clients and Servers Logical SIP entities are: User Agents User Agent Client (UAC): Initiates SIP requests User Agent Server (UAS): Returns SIP responses Network Servers Registrar: Accepts REGISTER requests from clients Proxy: Decides next hop and forwards request Redirect: Sends address of next hop back to client The different network server types may be collocated

21 SIP Addressing Uses Internet URLs Uniform Resource Locators Supports both Internet and PSTN addresses General form is To complete a call, needs to be resolved down to User@Host Examples: sip:ssavage@ucsd.edu sip:j.t. Kirk <kirk@starfleet.gov> sip: @ucsd.edu;user=phone sip:guest@

22 SIP Session Setup Example SIP User Agent Client INVITE SIP User Agent Server 200 OK ACK Media Stream BYE 200 OK host.cox.com sip.ucsd.edu

23 Proxy Server Example SIP User Agent Client SIP Proxy Server SIP User Agent Server INVITE INVITE 200 OK 200 OK ACK Media Stream BYE 200 OK host.cox.com server.cox.com sip.ucsd.edu

24 Redirect Server Example SIP User Agent Client SIP Redirect Server REGISTER SIP User Agent Server INVITE 200 OK 302 Moved ACK C 1 2 RS INVITE ssavage@ucsd.edu 3 UAS 180 Ringing 200 OK ACK Media Stream host.cox.com server.cox.com sip.ucsd.edu

25 SIP Requests SIP Requests (Messages) defined as: Method Method SP Request-URI SP SIP-Version CRLF (SP=Space, CRLF=Carriage Return and Line Feed) Example: INVITE SIP/2.0 Description INVITE A session is being requested to be setup using a specified media ACK Message from client to indicate that a successful response to an INVITE has been received OPTIONS A Query to a server about its capabilities BYE A call is being released by either party CANCEL Cancels any pending requests. Usually sent to a Proxy Server to cancel searches REGISTER Used by client to register a particular address with the SIP server

26 SIP Requests Example Required Headers (fields): INVITE SIP/2.0 Via: SIP/2.0/UDP host.cox.com:5060 From: Alan Johnston To: Stefan Savage Call-ID: CSeq: 1 INVITE }Uniquely identify this session request Via: Shows route taken by request. Call-ID: unique identifier generated by client. CSeq: Command Sequence number» generated by client» Incremented for each successive request

27 SIP Requests Example INVITE SIP/2.0 Via: SIP/2.0/UDP host.cox.com:5060 From: Alan Johnston To: Stefan Savage Call-ID: CSeq: 1 INVITE Contact: sip:joe.student@ucsd.edu Subject: Where are you these days? Content-Type: application/sdp Content-Length: 124 Field SDP data Descripton v=0 o=js IN IP4 host.cox.com s=let's Talk t=0 0 c=in IP m=audio RTP/AVP 0 3 Version v=0 Origin Session Name Times Connection Data o=<username> <session id> <version> <network type> <address type> <address> s=<session name> t=<start time> <stop time> c=<network type> <address type> <connection address> Media m=<media> <port> <transport> <media format list>

28 Lots more Gateways Extra services Emerging business

29 Token Buckets Common, simple descriptor Fill rate R tokens/sec Use tokens to send bits Average bandwidth is R bps Maximum burst is B bits Bucket size B tokens Sending drains tokens

30 Network Measurement The Internet is an artifact of sufficient complexity that we don t understand it Difficult to measure many things directly -> inference Too big to measure all of it No such thing as typical (sampling difficult) Constantly changing A whole community focused on these problems Packet characteristics: application usage; dynamic behavior of existing protocols; Path characteristics: delay, loss rate, queuing, bandwidth [end-to-end vs hop-by-hop] Network topology: connections between routers, between ISPs; graph analysis of Internet topologies Big challenge: no support for measurement

31 For example: How to measure one-way network path measurements (e.g. packet loss rate)? Critical for network performance analysis Requires measurement from both endpoints Remote hosts are indifferent to problem Unrealistic to deploy new software Remote hosts may block measurements

32 Key idea Exploit services needed by remote host Web, , news, file transfer, etc. All use Transmission Control Protocol (TCP)» Reliable, in-order, data transfer protocol Exploit standard protocol behavior TCP has rich specification Can be leveraged to perform measurements

33 Simple example: Deducing one-way loss rate Send TCP data packets to remote host What we know Number of data packets sent Number of acknowledgments received What we need to know How many data packets were received?» Extract from ACKs; TCP is a reliable protocol How many acknowledgments were sent?» Arrange that one ACK sent for each data packet

34 How TCP reveals packet loss Data packets ordered by sequence # s ACK packets specify next seq# expected Nothing lost Data lost ACK lost Time Data ACK 1 3

35 Loss deduction example Measurement 1 Loss deduction 4 Time 2 datasent = 3 ackreceived = Data ACK 2 2 datalost = 1 datareceived,acksent = 2 Loss data =1 - (datareceived/datasent) = 33% Loss ack =1 - (ackreceived/acksent) = 50% 5

36 Experimental finding: Packet loss is highly asymmetric 25 Popular Web servers Cumulative fraction To Web server From Web server Loss rate

37 What s s in a router? (how to make a very fast router?) Physical components One or more input interfaces that receive packets One or more output interfaces that transmit packets A chassis (box + power) to hold it all Functions Forward packets Drop packets (congestion, security, QoS) Delay packets (QoS) Transform packets? (Encapsulation, Tunneling)

38 What a router does: the normal case Receive incoming packet from link input interface Lookup packet destination in forwarding table (destination, output port(s)) Validate checksum, decrement ttl, update checksum Buffer packet in input queue Send packet to output interface (interfaces? Mcast) Buffer packet in output queue Send packet to output interface link

39 What a router looks like? Cisco 2500 Linksys DEFSR Capacity: <10Mbps Capacity: <10Mbps

40 What a router looks like (2) Cisco GSR Juniper M ft Capacity: 160Gb/s Power: 4.2kW 3ft Capacity: 80Gb/s Power: 2.6kW 2ft 2.5ft

41 What a router looks like (3) Pluris Teraplex 20 w/7 Racks Capacity: >1Tb/s Power: 45kW (~250 homes?) 1 room

42 High-performance routers Geared to core and distribution service needs Requirements: high speed & high density Why do we care? Moore s Law slower than link speed growth (and BW demand)» OC48c (2.5Gbps), now, 128ns/packet» OC192c (10Gbps), in deployment, 33ns/packet» OC768c (40Gbps), 2004, 8ns/packet Need high density/low power to manage POP complexity» $20-100k & 2-400W per port, 50% ports frequently for internal connectivity» DWDM can help with the former, but requires more interfaces

43 Functional architecture Firewall Reservation/ Admission Control Classification Rules Routing Protocols Routing Table Control Plane Complex Per-control action May be slow Packet Classification Forwarding Table Switching Output Scheduling Data plane Simple Per-packet Must be fast

44 Packet classification Forwarding Longest prefix match of destination against forwarding table Returns (output port, Next-hop MAC header) tuple Key issue: forwarding table growth QoS tagging Certain traffic tagged with higher priority Per flow (src ip, src port, dst ip, dst port), pre source or dest prefix, per protocol (Napster, etc ) Firewall rules Block access to TCP packets with dst port!= 80

45 Interconnect architecture Input & output connected via switch fabric Kinds of switch fabric Bus Crossbar Shared Memory Input Switch Output How to deal with transient contention? Input queuing Output queuing Combination

46 First Generation Routers Shared Bus(s) CPU Route Table Buffer Memory Single CPU and shared memory; All classification by main CPU Line Card Line Card Line Card MAC MAC MAC

47 Second Generation Routers CPU Route Table Shared Bus(s) Direct DMA on cache hit Cache of recent routes Line Card Buffers Line Card Buffers Line Card Buffers Forwarding Cache Forwarding Cache Forwarding Cache MAC MAC MAC

48 Third Generation Routers Switch Fabric Shared interconnect (usually crossbar) Centralized scheduler Full forwarding table in line card Fixed cells (why?) Line Card Buffers Forwarding Table MAC CPU Card CPU Routing Table Line Card Buffers Forwarding Table MAC

49 Output queuing Output interfaces buffer packets Input Switch Output Pro Simple algorithms Single congestion point Con N inputs may send to the same output Requires speedup of N

50 Input queuing Input interfaces buffer packets Input Switch Output Pro Single congestion point Simple to design algorithms Con Must implement flow control Low utilization due to Headof-Line (HoL) Blocking» Utili limited to 2-2^.5=58%

51 Head-of of-line Blocking

52 Virtual Output Queues (courtesy Nick McKeown)

53 IQ + Virtual Output Queuing Input interfaces buffer packets in per-output virtual queues Input Switch Output Pro Con Solves blocking problem More resources per port Complex arbiter Still limited by input/output contention (scheduler). RR: 1=1/e = 63%

54 Switch scheduling Problem Match inputs and outputs Resolve contentions No output packet drops Maximize throughput Do it in constant time Many algorithms for uniform traffic assumption E.g. TDM, Maximum size bipartite match Approximate answers (e.g. islip, submaximal match) Recent result (Dai et al, 2000) Maximal size matching + speedup of two guarantees 100% utilization for most traffic assumptions

55 Modern router IQ + VoQ + OQ Speedup of 2 Central scheduler Input Switch Output Fixed-sized internal cells Pro Can achieve utilization of ~1 Can scale to multiple Tb/s Con Multiple congestion points Complexity

56 Typical function breakdown Input interface Switch Forwarding Virtual output queuing Scheduling input interface requests to output interfaces Multicast scheduling Output interface Queue packets for transmission Classification Buffer management (which pkt to drop) Scheduling (which pkt to send from buffer)

57 Next bottlenecks Buffering at high speed SRAM density too low for BW*D of 40Gbps link DRAM too slow SRAM memory management as cache for DRAM Scheduler and arbiter overhead Limits size of switch and link BW Two-state switch (Chang et al, 2000); no scheduler High density (100 s-1000 s of line cards) Physical distance to support density; electrical links degrade Optical links; optical cross connect (MEMs,tunable lasers) Time to market, Power/Heat

58 Next time I will see you folks at office hours, the review session or the final Good luck, its been a pleasure.