Gigabit Ethernet

Conventional infrastructure trying to solve yesterdays problem (the 56k modem). Not looking forward far enough.
Building gigabit Ethernet native mode fiber rings for metropolitan area networks.

Broadband == isochronous information + data (integrated services)
$1 trillion in next 9 years

wired copper (xDSL, cable, ISDN, ATM) $200 B
fiber
satellite
wireless

MAN

12 miles
lots built already

RAN

regional
60 miles
a few in US as of 1999

New media, new infrastructure: the medium is immaterial, cost-wise

Distance, latency & jitter, isochronous traffic, capacity, resiliency - all addressed with gigabit Ethernet

Experiment

wire with switches, San Fran to NY, repeater switches every 37km, fiber
25us latency per switch + propogation delay == 46 ms total latency (very good)

L2: Ethernet - in the standards, in the switches
L3: IPv4 using those old fields, now possible to do routing at wire-speed (routing at 1 G bps), all legacy devices preserve this info
L4: standards in the works

Pacific northwest
13 agencies, 120 sites
voice, video, data
dark fiber by utility company
5 bidders of ATM, Packet Engines did gigabit Ethernet
10 year lifespan, no early upgrades, minimum traffic spec per site

ATM on OC-3 was about same cost as Gigabit Ether, but the performance is X times bigger

private rings for each agency, backbone ring common to all agencies

allows for staged implementations
rings give redundancy - cheaper/more flexible than SONET
routers adapt to ring destruction - better than physical layer redundancy

voice/IP gateway connects to ring at 10Mbps

compared to T1 line you break even in 6-16 months, after that you save money every month
greater initial hardware cost
no monthly access fee to phone company

project saves school district $100,000 / year

LANs came from PC industry, not PSTN. All the WAN stuff was done for voice, adapated for data.
Moore's Law not slowing down, extrapolate the CPU speed out to 2009 (14th generation Pentium)

 

	8088	PII	P14
CPU speed	4.77 MHz	350 M	10 GHz
bus speed	4.77 MHz	66 M	600 MHz
bus width	8 bit	64	256 bits

Therefore gigabit Ethernet will be on the desktop in the next 2-3 years

10 G Ethernet by 2002 reasonable prices

Trends and Technologies in High-Speed Switching

Layer 2 - switching, Layer 3 - routing
Switches are

signalling controlled
acts on local information, doesn't have global network view
hardware oriented, versus software for routers
more autonomous behavior - less configuration, simpler

WDM, SDH, ATM
10 G bps coming soon

key bottleneck
header processing
buffer management, qos handling
billing/charging
flow control, network protection

processing of packets this is the limiting factor
memory speed isn't increasing
board/chip interconnect are problems - moving from 128 bit wide bus to 1 G serial bus

network processor: RISC core + network tasks (table search, CRC error, etc)
analogous to signal processors
also good for NICs

Gigabit Ethernet

8002.3z

CSMA/CD and Full Duplex
fiber and copper
Fiber Channel 8 bit -> 10 bit encoding

pick the the symbols which are best to transmit (no DC, fewer transitions)

Fast convergence/adoption in the marketplace (as fast or more than 100 M)

Jumbo Frames (9k bytes)
VLAN

virtual LANs
since you have 10 -> 1000 G

Protocol stack parameter modifications

old protocols don't scale to 1G

Multimode

50/125 and 62/125 (inner core/outer fiber diamter, in microns)
typical of what is installed
850nm

200 - 500 m, cost

1300nm

550m, distance

Singlemode

9/125
MANs are possible
5km loing wave optics
10km with "hot optics"
80km with "proprietary optics"

STP/Coax

25m over 2 150 Ohm pairs, or 2 75 Ohm coax
relatively low cost and simple
useful for racks, single room
1250 baud, 8b/10b encoding

UTP Cat 5, 1000baseT, 802.3ab

100m all 4 pair
complex: requires forward error correction
250 Mbit, 125 M baud, 4 pair

Pure scaling

slot time is 512ns, max diameter is 28-43 m
hard to achieve efficiency at half duplex

Increased slot-time

to 4 mirosecond, 200m collision domain
packet bursting needed to increase efficiency

Full Duplex

the best solution - forget CSMA/CD
capacity = 2 * half duplex
not Ethenet!
wire-speed now realistic
limited only by transmission characteristics

DMD, differential mode delay

square pulses become Gaussian pulses due to differential propogation modes, speeds, attenuation
affects some modes more than others - not smooth as previously thought
reduces distance due to jitter and inter-symbol interference
single mode (skinny) fibers fix this problem by only allowing one mode through

Already more common than CSMA/CS in 1998

Disks and Networks

nearly same throughput
Fiber channel storage
Server area networks
Storage networks, U-SCSI and GBE
small packet size in GBE is major difference

Metropolitan Area Networks

best price/performance for GBE
10000 of users at 100 k bps each
100s of concurrent MPEG-2 video streams
multicast video is possible
cheap optical fiber adaptors on the market

Gigabit adapters and bus bandwidth

need more than PCI
proprietary buses
tons of interrupts, even on a fast machine due to many short frames - argument for better kernel handling

Redundant physical media fault tolerance

in siwtches, but not in NICs
may never happen

Jumbo frames
Cut through

efficient lock management

high speed reduces need
no specific characteristic

Hunt groups

aggregated links

10 G

possibly 4, 2.5 G with WDM

Switched GBE cost 1999

$1000 per port
single ASIC for switch and NIC functionality

1550nM GBICs

100km distance