Administrivia Fishnet Assignment #1 P561: Network Systems � � Due next week (week 3), start of class Week 2: Local Area Networks � � Electronic turnin � � No class trawler (do that for Fishnet #2) Tom Anderson � Homework #1 Ratul Mahajan � � � On web site � � Due two weeks (week 4), start of class TA: Colin Dixon � Next week: Internetworking, broadcast from MSR 2 Q&A from last time Network Building Blocks Links – carry information (bits) How far can an optical link go without a repeater? � � Wire, optics or wireless � � About 20 km in practice � � Point to point or broadcast � � 10 terabits/100 km in prototypes Switches/Routers -- move bits between links � � packet or circuit switching Why do they call it MIMO beamforming? Host – communication endpoint � � Can independently control the phase and amplitude of � � computer, PDA, toaster, ... each antenna, which affects the receiver power. Network -- delivers messages between hosts over a collection of links and switches 3 Internet Design Goals Network Sharing Effective multiplexing of existing networks Networks are shared among users � � This is an important benefit of building them � � multiplexing = sharing � � using store & forward packet switching Problem: How to multiplex (share) a link among multiple Survivability in the face of failure users? � � Communication must continue despite loss of equipment Well, we could statically partition the link: Heterogeneity � � Frequency Division Multiplexing (FDM) � � In networks and applications � � (Synchronous) Time Division Multiplexing (TDM, STDM) Distributed management 5 6 1
Frequency Division Multiplexing Time Division Multiplexing Timeslice given frequency band between users Simultaneous transmission in different frequency bands � � Digital: used extensively inside the telephone network � � T1 (1.5Mbps) is 24 x 8 bits/125us; also E1 (2Mbps, 32 � � Analog: Radio/TV, AMPS cell phones (800MHz) slots) � � Also called Wavelength DM (WDM) for fiber timeslot phone call time � freq � “Speaking at different times” guard bands Advantage: lower delay “Speaking at different pitches” Disadvantage: synchronization 7 8 Statistical Multiplexing Example One user sends at 1 Mbps and is idle 90% of the Static partitioning schemes are not suited to data time. communications because peak rate >> average � � 10 Mbps channel; 10 users if statically allocated rate. Prob Prob 2 users 10 users If we share on demand we can support more users � � Based on the statistics of their transmissions � � Occasionally we might be oversubscribed � � This is called statistical multiplexing 0 1 2 Mbps 0 1 … 10 Mbps Statistical multiplexing is heavily used in data What are the likely loads if we share on demand? networks 9 10 Example continued Workload Questions For 10 users, Prob(need 10 Mbps) = 10 -10 How bursty is the data traffic to/from a single node? Not likely! So keep adding users … � � Self-similar at many time scales For 35 users, Prob(>10 active users) = 0.17%, which is acceptably low How bursty is the data traffic to/from a campus? We can support three times as many users! How bursty is the data traffic in the core of the But: there is an important caveat here … Internet? � � Elephants and mice 11 12 2
ALOHA Carrier Sense Multiple Access We can do better by listening before we send Packet radio network in Hawaii, 1970s (CSMA) Wanted distributed allocation � � good defense against collisions if “a” is small � � no special channels or single point of failure Aloha protocol: � � Just send when you have data! (wire) X � � There will be some collisions of course … collision � � Throw away garbled frames at receiver (using CRC); A B sender will time out and retransmit Simple, decentralized and works well for low load “a”: number of packets that fit on the wire � � What happens when load increases? � � bandwidth * delay / packet size � � Small for LANs; large for satellite What if the Channel is Busy? CSMA with Collision Detection 1-persistent CSMA Even with CSMA there can still be collisions. Why? � � Wait until idle then go for it � � Blocked senders can queue up and collide p-persistent CSMA Time for B to detect A’s transmission � � If idle send with prob p in each time slot until done (wire) X collision � � Choose p so p * # senders < 1; how do you know p? A B non-persistent CSMA � � Wait a random time and try again For wired media we can detect all collisions and abort (CSMA/CD): � � Better when loaded, but larger delay when unloaded � � Requires a minimum frame size (“acquiring the medium”) � � B must continue sending (“jam”) until A detects collision 15 16 Classic Ethernet Ethernet Frames (Classic) IEEE 802.3 standard wired LAN (1-persistent CSMA/CD) Classic Ethernet: 10 Mbps over coaxial cable � � baseband signals, Manchester encoding, preamble, 32 bit CRC Preamble (8) Source (6) Dest (6) Len (2) Payload (var) Pad (var) CRC (4) (wire) Min frame 64 bytes, max 1500 bytes nodes CSMA/CD jam period is 48 bits Newer versions are much faster Max length 2.5km, max between stations 500m � � Fast (100 Mbps), 1 Gb, 10Gb (repeaters) Hub or Modern equipment isn’t one long wire Addresses unique per adaptor; globally assigned Switch � � hubs and switches Broadcast media: � � ARP, multicast, promiscuous mode monitoring 17 3
Binary Exponential Backoff Ethernet Capture Build on 1-persistent CSMA/CD Randomized access scheme is not fair On collision: jam and exponential backoff � � Jamming: send 48 bit sequence to ensure collision Stations A and B always have data to send detection � � They will collide at some time Backoff: � � Suppose A wins and sends, while B backs off � � First collision: wait 0 or 1 frame times at random and � � Next time they collide and B’s chances of winning are retry halved! � � Second time: wait 0, 1, 2, or 3 frame times � � Nth time (N<=10): wait 0, 1, …, 2 N -1 times � � Max wait 1023 frames, give up after 16 attempts � � Scheme balances average wait with load 19 20 Ethernet Performance Why Did Ethernet Win? Much better than Aloha or CSMA! � � Works very well in practice Reliablity � � Token ring failure mode -- network unusable Source of protocol inefficiency: collisions � � Ethernet failure mode -- node detached Cost � � More efficient to send larger frames • � Acquire the medium and send lots of data � � Passive tap cheaper to build than active forwarder � � Less efficient as the network grows in terms of frames � � Volume => lower cost => volume => lower cost … • � recall “a” = delay / (frame size * transmission rate) Scalability • � “a” grows as the path gets longer (satellite) � � Repeater: copy all packets across two segments • � “a” grows as the bit rates increase (Fast, Gigabit Ethernet) � � Bridge: selectively repeat packets across two segs � � Switch: bridge k segments; Hub: repeater for k segs 21 Switched Ethernet Scaling Build larger networks out of small building blocks What if # of nodes > degree of one switch? Redundancy for higher availability � � What does a data center network look like? Simple case: # of nodes < degree of switch 23 24 4
Fat Trees Internet PoPs Bisection bandwidth: the minimum bandwidth PoP = Point of Presence between any equal partitioning of the nodes � � Use redundancy at each level to mask failures � � Important if network communication is all to all 25 26 Internet PoPs Inside a Switch PoP = Point of Presence If switch degree is small enough, use a crossbar � � Use redundancy at each level to mask failures � � Need buffering at the inputs � � Performance degrades (badly) with head of line blocking 27 28 Inside a Switch Wireless Communication What if you want to build a wider switch? Wireless is more complicated than wired … Cannot detect collisions 1. � � � Transmitter swamps co-located receiver Different transmitters have different coverage 2. � areas � � Asymmetries lead to hidden/exposed terminal problems 29 5
Recommend
More recommend