Chapter Peer-to-Peer Protocols and Data Link Layer PART I: Peer-to-Peer Protocols Peer-to-Peer Protocols and Service Models ARQ Protocols and Reliable Data Transfer Flow Control Timing Recovery TCP Reliable Stream Service & Flow Control Chapter Peer-to-Peer Protocols and Data Link Layer PART II: Data Link Controls Framing Point-to-Point Protocol High-Level Data Link Control Link Sharing Using Statistical Multiplexing Chapter Overview z Peer-to-Peer protocols: many protocols involve the interaction between two peers z z z z Service Models are discussed & examples given Detailed discussion of ARQ provides example of development of peer-to-peer protocols Flow control, TCP reliable stream, and timing recovery Data Link Layer z z z Framing PPP & HDLC protocols Statistical multiplexing for link sharing Chapter Peer-to-Peer Protocols and Data Link Layer Peer-to-Peer Protocols and Service Models zzz zzz Peer-to-Peer Protocols n + peer process SDU PDU z Layer-(n+1) peer calls layer-n and passes Service Data Units (SDUs) for transfer z Layer-n peers exchange Protocol Data Units (PDUs) to effect transfer z Layer-n delivers SDUs to destination layer-(n+1) peer n peer process n – peer process zzz zzz n – peer process Peer-to-Peer processes execute layer-n protocol to provide service to layer-(n+1) n + peer process SDU n peer process z Service Models z z The service model specifies the information transfer service layer-n provides to layer-(n+1) The most important distinction is whether the service is: z z z Connection-oriented Connectionless Service model possible features: z z z z z Arbitrary message size or structure Sequencing and Reliability Timing, Pacing, and Flow control Multiplexing Privacy, integrity, and authentication Connection-Oriented Transfer Service z Connection Establishment z z z Message transfer phase z z z Connection must be established between layer-(n+1) peers Layer-n protocol must: Set initial parameters, e.g sequence numbers; and Allocate resources, e.g buffers Exchange of SDUs Disconnect phase Example: TCP, PPP n + peer process send SDU n + peer process receive Layer n connection-oriented service SDU Connectionless Transfer Service z z z z z No Connection setup, simply send SDU Each message send independently Must provide all address information per message Simple & quick Example: UDP, IP n + peer process send SDU n + peer process receive Layer n connectionless service Message Size and Structure z What message size and structure will a service model accept? z z z z Different services impose restrictions on size & structure of data it will transfer Single bit? Block of bytes? Byte stream? Ex: Transfer of voice mail = long message Ex: Transfer of voice call = byte stream voice mail= message = entire sequence of speech samples (a) call = sequence of 1-byte messages (b) Segmentation & Blocking z z z To accommodate arbitrary message size, a layer may have to deal with messages that are too long or too short for its protocol Segmentation & Reassembly: a layer breaks long messages into smaller blocks and reassembles these at the destination Blocking & Unblocking: a layer combines small messages into bigger blocks prior to transfer long message or more blocks or more short messages block M/M/1/K Performance Results (From Appendix A) Probability of Overflow: (1 − ρ ) ρ K Ploss = K +1 1− ρ Average Total Packet Delay: ρ ( K + 1) ρ − E[ N ] = K +1 1− ρ 1− ρ E[ N ] E[T ] = λ (1 − PK ) K +1 normalized avg delay E[T]/E[X] M/M/1/10 10 z z loss probability los s probability 0.2 0.4 0.6 0.8 1.2 1.4 1.6 1.8 2.2 2.4 2.6 2.8 load 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 z z 0.3 0.6 0.9 1.2 1.5 lo a d 1.8 2.1 2.4 2.7 z Maximum 10 packets allowed in system Minimum delay is service time Maximum delay is 10 service times At 70% load delay & loss begin increasing What if we add more buffers? M/M/1 Queue Poisson Arrivals rate λ Infinite buffer Exponential service time with rate μ Unlimited number of customers allowed in system z z z z z Pb=0 since customers are never blocked Average Time in system E[T] = E[W] + E[X] When λ μ customers arrive faster than they can be processed and queue grows without bound (unstable) Avg Delay in M/M/1 & M/D/1 10 no rmalized avg delay normalized average delay M/M/1 constant service time M/D/1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99 lo a d E[TM ] = ⎡ ρ ⎤1 1⎡ ρ ⎤ ⎡ ⎤1 = = ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ + λ ⎣1 − ρ ⎦ ⎣1 − ρ ⎦ μ ⎣1 − ρ ⎦ μ μ ⎡ ρ ⎤1 ⎡ ρ ⎤1 E[T D ] = ⎢1 + ⎥ + ⎥ =⎢ ⎣ 2(1 − ρ ) ⎦ μ ⎣ 2(1 − ρ ) ⎦ μ μ for M/M/1 model for M/D/1 system Effect of Scale z z z C = 100,000 bps Exp Dist with Avg Packet Length: 10,000 bits Service Time: X=0.1 second Arrival Rate: 7.5 pkts/sec z Load: ρ=0.75 z Mean Delay: E[T] = 0.1/(1-.75) = 0.4 sec z C = 10,000,000 bps z Exp Dist with Avg Packet Length: 10,000 bits z Service Time: X=0.001 second z Arrival Rate: 750 pkts/sec z Load: ρ=0.75 z Mean Delay: z E[T] = 0.001/(1-.75) = 0.004 sec Reduction by factor of 100 z Aggregation of flows can improve Delay & Loss Performance Example: Header overhead & Goodput z z z Let R=64 kbps Assume IP+TCP header = 40 bytes Assume constant packets of total length z z z z z z L= 200, 400, 800, 1200 bytes Find avg delay vs goodput (information transmitted excluding header overhead) Service rate μ = 64000/8L packets/second Total load ρ = λ 64000/8L Goodput = λ packets/sec x 8(L-40) bits/packet Max Goodput = (1-40/L)64000 bps Header overhead limits maximum goodput Average Delay (seconds) 0.6 0.5 L=1200 0.4 L=800 0.3 L=400 0.2 0.1 L=200 0 8000 16000 24000 32000 40000 48000 56000 64000 Goodput (bits/second) Burst Multiplexing / Speech Interpolation Many Voice Calls Fewer Trunks Part of this burst is lost z z z z z Voice active < 40% time No buffering, on-the-fly switch bursts to available trunks Can handle to times as many calls Tradeoff: Trunk Utilization vs Speech Loss z Fractional Speech Loss: fraction of active speech lost Demand Characteristics z Talkspurt and Silence Duration Statistics z Proportion of time speaker active/idle Speech Loss vs Trunks trunks 10 12 14 16 18 20 22 24 speech loss 0.1 0.01 Typical requirement 48 0.001 24 32 40 # connections ⎛n⎞ k (k − m)⎜ ⎟ p (1 − p ) n − k ⎝k⎠ n! ⎛n⎞ k = m +1 speech loss = where ⎜ ⎟ = ! ( )! np k n − k k ⎝ ⎠ n ∑ Effect of Scale z z Larger flows lead to better performance Multiplexing Gain = # speakers / # trunks Trunks required for 1% speech loss Speakers Trunks Multiplexing Gain Utilization 24 13 1.85 0.74 32 16 2.00 0.80 40 20 2.00 0.80 48 23 2.09 0.83 Packet Speech Multiplexing Many voice A3 terminals B3 generating voice packets A2 A1 B2 B1 C3 C2 C1 D3 D2 D1 Buffer B3 C3 A2 D2 C2 B1 C1 D1 A1 Buffer overflow B2 z z z z z Digital speech carried by fixed-length packets No packets when speaker silent Synchronous packets when speaker active Buffer packets & transmit over shared high-speed line Tradeoffs: Utilization vs Delay/Jitter & Loss Packet Switching of Voice Sent Received z z z t Packetization delay: time for speech samples to fill a packet Jitter: variable inter-packet arrivals at destination Playback strategies required to compensate for jitter/loss z Flexible delay inserted to produce fixed end-to-end delay z Need buffer overflow/underflow countermeasures z Need clock recovery algorithm t Chapter Peer-to-Peer Protocols and Data Link Layer ARQ Efficiency Calculations Stop & Wait Performance successful transmission i – unsuccessful transmissions ∞ E [t t o t a l ] = t + ∑ (i − 1)t o ut P[nt = i ] i =1 =t ∞ + ∑ (i − 1)t o ut (1 − Pf )i −1 Pf i =1 =t + t o u t Pf − Pf =t 1 − Pf Efficiency: n f − no η SW E[ttotal ] = = (1 − Pf ) R 1− na 1+ + nf no nf 2(t prop + t proc ) R nf = (1 − Pf )η Go-Back-N Performance i – unsuccessful transmissions successful transmission ∞ E[ttotal ] = t f + ∑ (i − 1)Ws t f P[nt = i ] i =1 ∞ = t f + Ws t f ∑ (i − 1)(1 − Pf ) i −1 Pf i =1 = tf + Ws t f Pf − Pf Efficiency: = tf n f − no η GBN = + (Ws − 1) Pf − Pf no 1− nf E[ttotal ] = (1 − Pf ) R + (Ws − 1) Pf