EECS 122, Lecture 7. Kevin Fall Jean Walrand

Transcription

1 EECS 122, Lecture 7 Kevin Fall kfall@cs.berkeley.edu Jean Walrand wlr@eecs.berkeley.edu

2 : Outline Typical Setup Names Physical Layer Frame Fast Ethernet; Gigabit Ethernet 10Base5 Efficiency of CSMA/CD Perspective EECS Fall & Walrand 198

3 Typical Setup EECS Fall & Walrand 199

4 Names Names of form [rate][modulation][media or distance] Examples: n 10Base2 (10Mb/s, baseband, small coax, 200m) n 10Base5 (10Mb/s, baseband, coax, 500m) n n n 10Base-T (10Mb/s, baseband, twisted pair) 100Base-TX (100Mb/s, baseband, 2 pair) 100Base-FX (100Mb/s, baseband, fiber) EECS Fall & Walrand 200

5 Physical Layer EECS Fall & Walrand 201

6 Physical Layer EECS Fall & Walrand 202

7 Physical Layer Hub: Single Collision Domain MAC Protocol: Wait until silent (carrier sense) Transmit CSMA/CD If collision, wait random time & repeat EECS Fall & Walrand 203

8 Physical Layer Switch: No Collisions Multiple transmissions are possible Switch stores packets that wait for same output EECS Fall & Walrand 204

9 Frame 7 byte preamble: alternating 1/0 combination producing 10Mhz square wave [@ 10Mbps] for 5.6 µsec; used for receiver synchronization 1 byte SFD (start of frame delimiter) EECS Fall & Walrand 205

10 Frame Length/Type field: Type (Ethernet) indicates type of data contained in payload issue: what is the length? Length field (802.3) type info follows frame header So, is it the type or length? Ethernet : types have values above 2048 (RFC894 for IP) 802.3: length (RFC1042 for IP) If length, next headers are LLC & SNAP (for IP) LLC (3 bytes): DSAP, SSAP, CTL SNAP (5 bytes): org code, type (above) EECS Fall & Walrand 206

11 IEEE 802.3u 100 Mb/s Fast Ethernet (1995) adds: n 10x speed increase (100m max cable length retains min 64 byte frames) n replace Manchester with 4B/5B (from FDDI) n full-duplex operation using switches n speed & duplex auto-negotiation EECS Fall & Walrand 207

12 IEEE 802.3{z,ab} 1000 Mb/s Gigabit Ethernet (1998,9) adds: n 100x speed increase n carrier extension (invisible padding...) n packet bursting EECS Fall & Walrand 208

13 10base5 - Definition Will discuss classical Ethernet primarily Single segments up to 500m; with up to 4 repeaters gives 2500m max length Baseband signals broadcast, Manchester encoding, 32-bit CRC for error detection Max 100 stations/segment, 1024 stations/ethernet EECS Fall & Walrand 209

14 10base5 Collision Detection CD circuit operates by looking for voltage exceeding a transmitted voltage Want to ensure that a station does not complete transmission prior to 1st bit arriving at farthest-away station Time to CD can thus take up to 2x{max prop. delay} A B time EECS Fall & Walrand 210 CD

15 10base5 minimum frame size Speed of light is about 4µs/km in copper So, max Ethernet signal prop time is about 10 µsec, or 20µsec RTT With repeaters, etc requires 51usec, corresponding to 512 bit-times Thus, minimum frame size is 512 bits (64 bytes); also called slot time EECS Fall & Walrand 211

16 10base5 max. frame size 1500 byte limitation on maximum frame transmission size Later we will call this the MTU limits maximum buffers at receiver allows for other stations to send n also requires 96 bit Inter-Packet-Gap (IPG) EECS Fall & Walrand 212

17 10base5 - Transmitter When ready & line idle, await IPG (96 bit times) and send while listening (CD) If CD true, send max 48-bit jamming sequence and do exponential backoff Jamming sequence used to inform all stations that a collision has occurred EECS Fall & Walrand 213

18 10base5 Exponential Backoff For retransmission N (1<=N<=10) n choose k at random on U(0..2^N-1) n wait k * (51.2µsec) to retransmit n send on idle; repeat on another collision n for (11<=N<=15), use U( ) n if N = 16, drop frame Longer wait implies lower priority (strategy is not fair ) EECS Fall & Walrand 214

19 10base5 Capture Effect Given two stations A & B, unfair strategy can cause A to continue to win Assume A & B always ready to send: n if busy, both wait, send and collide n suppose A wins, B backs off n next time, B s chances of winning are halved EECS Fall & Walrand 215

20 Efficiency of CSMA/CD Typical Sequence of Events: Average = 5T T = max.prop.time between 2 nodes Average = L/R seconds L = average packet length Wait Random Successful Transmission Time Collision L/R Efficiency = = 1/(1 + 5a) Start L/R + 5T Transmitting a = T/(L/R) = RT/L EECS Fall & Walrand 216

21 Efficiency of CSMA/CD a impacts what happens during simultaneous transmission: n a small early collision detection Efficient a = RT/L eff = 1/(1 + 5a) n a large late detection Inefficient Example 1: 10Mbps, 1000m => T = (1km)(4µs/km) = 4µs; RT = 400 bits L = 4000 bits, say 5a = 2000/4000 = 0.5 => efficiency = 66% Example 2: 1Gbps, 200m => T = (0.2km)((4µs/km) = 0.8µs; RT = 800 bits L = 4000 bits; 5a = 4000/4000 = 1 => efficiency = 50% EECS Fall & Walrand 217

22 Efficiency of CSMA/CD - Analysis Model: Slot = 2T N stations compete by transmitting with probability p, independently If success => transmit L bits If failure (idle or collision), try next slot P(success) = P(exactly 1 out of N transmits) = Np(1 p) N-1 Indeed: N possibilities of station that transmits P(one given station transmits, others do not) = p(1 p) N-1 EECS Fall & Walrand 218

23 Efficiency of CSMA/CD - Analysis Average = A Slot = 2T success P(success) = Np(1 p) N-1 Assume backoff algorithm results in best p = 1/N => P(success) 1/e = 0.36 Average time until success: A = 0.36x x(2T + A) => A = 1.28T/0.36 = 3.5T In practice, backoff not quite optimal => 5T EECS Fall & Walrand 219

24 Addressing 48 bit Ethernet/MAC/Hardware Addresses Prefix assigned per-vendor by IEEE Unique per-adapter, burned in ID PROM Multicast & Broadcast (all 1 s) addresses Many adapters support promiscuous mode EECS Fall & Walrand 220

25 Addressing: Multicast Each vendor assignment supports 2^24 individual and group (multicast) addresses Each adapter supports multiple group subscriptions n usually implemented as hash table n thus, software may have to filter at higher layer EECS Fall & Walrand 221

26 Perspective Ethernet is wildly successful, partly due to low cost (compare with FDDI or Token Ring--- see text book) Some issues: n nondeterministic service n no priorities n min frame size may be large EECS Fall & Walrand 222