Intel 2 Gbps pure silicon based optical modulator
smashes previous optical switching record
of 20 Mbps held by Graham Reed, University
of Surrey, UK.
OXC "fastest" optical cross connects promise to be the
fastest switching technology between
high speed fibers, but it is not yet
clear how well they integrate into IP
over optical networks.
MPLS multiprotocol label switching is an
effort to traffic engineer high
volume flows and may ease the integration
of ATM, VCs and IP. It may be more
efficient to utilize load sensitive IP
streams if they could be implemented
without instabilities for example by
optimising OSPF weights or e.g. using an
ensemble Riccati state controller.
Avici fast IP-in-silicon fast IP router.
BBN Labs 50 Gbps research effort based on the Butterfly
switching matrix, ARPA funded, Craig
Partridge involved, forwards 6 to 42
million packets per second depending
on configuration. This was one of the
earliest multigigabit routers built.
British Telecom 100 Gbps research effort, pure optical routing,
read header in optical IC and marked
packets are optically switched into
another fiber, 100 Gbps expected for
the first light box.
Broadcom 9 Gbps world's first single chip switch with
routing (L3) and quality of service. The
Strataswitch BCM5600 (CMOS) can filter up
to 6.6 Mpps and supports 24 10/100
interfaces and two 100/1000 interfaces
which can all run at wire speed. The
chip can run data, voice and video
Caspian Networks very fast next generation scalability, control and
management of ISP's IP optical networks,
Lawrence Roberts involved.
Chiaro Networks very fast developped optical switching, eliminating
the capacity bottleneck (due to electronics
using slow large electrons instead of light
and fast photons) at the carrier's end of
Cisco 7600 OSR 256 Gbps The 7600 Optical Services Router (OSR)
has a parallel eXpress forwarding engine,
non ASIC based, which can handle up to 30
million packets per seconds. A variety
of configurations with gigabit ethernet
and ATM interfaces etc are available.
Introduced in February 2001 and put on
hold in April 2001.
i-data/Exbit 32 Gbps The Heathrow 16-port-switch-on-a-chip can
handle 10/100/1000 speeds on each port,
24 million packets per second and has an
internal engine up to 32 Gbps.
Lucent gigabits The Packet Star terabit switch is an
IP-in-silicon router using PGAs without
the traditional CPU and software lookups
on each incoming packet. It using a
clever packet queuing scheme and
includes QoS (ATM).
Moby Router 350 Gbps research effort based on silicon, parallel
Pluris architectures, Vadim Antonov is founder
of Pluris, 2 Tbps targeted. IP routing in
silicon, not the traditional CPU and
software approach. Pluris core technology
works by splitting aggregated streams into
parallel flows, like 'colored liquids'.
Expected to have advantages over simple IP
best efforts routing on the one side and
complicated virtual circuit management on
the other. (When circuits aren't over-
loaded traffic engineering has zero
benefits. The TE apostles promise more
"miracles" for unfilled pipes.)
Nexabit/Lucent acq. gigabits developping the capability to route up
to 10 billion bits per second real world
on each interface.
Procket fast the next generation router, faster than
the GSR 12000 or M160, using MPLS.
Rapid City Communic. 7 Mpps startup in Mountain View, Gigabit ethernet
routing switches can route 7 million
packets per second, uses ASICs, expected
to ship end of 1997, was purchased by Bay
Networks, now Nortel Networks.
Rumor by Bill Manning 600 Mbps 'thingie with PCI interface', box needed
to plug this into, to get a reasonable
ExchangeNG(tm) box, hopefully with multiple
Or simply four gigabit PCI cards,
although it's nontrivial to get this box
to work with present day hardware, due to
memory bandwidth constraints, which is
only 4.2 Gbps in a PC and requires 2 Gbps
for each PCI card, in and out; packet
header lookup and route processing etc.;
but it could become a future alternative
to custom hardware solutions; before
$200 four port gigabit ASIC boxes take
SNRC SupraNodes 100 Tb/s SupraNodes scalable to 100 Tb/s and
beyond are investigated using scalable
architectures and emerging optical &
electrical technologies at the SNRC
(Stanford Network Research Center)
here in town.
Vitesse 10 Gbps Pacemaker 10 chip planned for release in
Q4 2000. Full OC-192 speed switching and
packet processing engine.
Ascend GRF 400 16 Gbps the Gigarouter is a gallium arsenide
or GRF 1600 280 Mpps based non-blocking crosspoint switch,
16x16 matrix with 256 paths, used at
UUnet/Worldcom, on the NSF funded
vBNS OC-3c 155 Mbps backbone; 16 Gbps
advertised aggregate bandwidth; can hold
150,000+ routing table entries; does not
use routing table caches to reduce
flapping; available with HiPPI/ATM/OC3c/
FDDI interfaces, available with 4 slot
(GRF 400) or 16 slot (GRF 1600) rack,
uses gateD version, 'magic' runs below
socket level, developped by Netstar, an
Ascend division in Minnesota; won router
of the year award '97. Lucent bought
Bay Networks BCN(R) 700 Mbps Backbone Concentrator Node multiprotocol
>1 Mpps routers use a Symmetric MultiProcessing
(SMP) architecture (Bay Networks claims
their routers (BCN, BLN, ASN) are the only
ones which do; will be used soon on the
ANS DS3 backbone (replacing IBM RS6000s);
up to 13 FDDI interfaces; up to 104 LAN/WAN
interfaces; system performance scales to
800,000 to 1 million forwarded packets
per second (pps); runs SMP over up to 13
68060s with up to 416 MB DRAM, part of
which is used to mirror global memory to
local memory. ATM Routing Module can
handle wire speed OC-3.
Bay Networks BLN 330,000 pps Backbone Link Nodes can forward 330,000
pps, with SMP over up to 4 68060s and
128 MB RAM, part of which is used to
mirror global memory to local nodes.
Please supply better real world figures
if you have them.
Bay Networks ASN 50,000 pps Access Stack Nodes are modular stackable,
50,000 pps (1 ASN) up to 200,000 pps
(4 ASNs), based on 68040 CPUs with SMP
over up to 4 CPUs and up to 128MB DRAM,
part of which is used for mirroring.
Bay Networks AN 7000 pps Access Node can forward 7000 pps.
Cisco CRS-1 92 Tbps The Carrier Routing System 1 uses up to
40 Gbps fiber ports, up to 1.2 Gbps per
box. 72 boxes can be connected resulting
in an aggregate bandwidth of 92 Tbps.
Cisco GSR up to 80 Gbps The Gigabit Switch Router is a crosspoint
switch and router available with 4, 8 and
12 slots. Supports IP over SONET as well
as ATM. Connects directly into fiber
infrastructure with OC-3 or OC-12 speeds,
lateron OC-48. During development it
was known as the BFR (Big Fast Router).
GSR 12012 15-60 Gbps 12 slots, interface cards with 4 x OC-3
GSR 12008 10-40 Gbps 8 slots or 1 x OC-12
GSR 12004 5 Gbps 4 slots
Probably the best architecture: IP over
SONET on long distance links, Gigabit
ethernet on the LAN and in between ATM
may not be all that bad for citywide
connections. Can do 2.5 billion bits per
seconds full duplex real world traffic
on each interface.
Cisco 7500 series 160 Mbps popular at MAEs, FIXs and backbones,
>1 Mpps dual MIPS processors, supports many
protocols and options suitable for NSPs,
about 160 Mbps real world performance
(in Cisco HSSI/FDDI testbed, possibly
limited by slow test equipment); they
are widely used on present day backbones
and can barely handle present day traffic;
can do about 300,000 packets per second;
and up to 1 Mpps with recent upgrades;
performance is negatively impacted by
route updating; route damping can
alleviate it by reducing the load on the
processors. SSH available in 12.xT IOS.
Cisco 7200 series > 16 Mpbs popular with ISPs, 8 Mbps on one, maybe
two, ports of a 4T card.
Cisco 4700 series > 10 Mbps recommended for full BGP4 routing with
about 34,000 routes and using 32 MB RAM;
100 MHz 64bit IDT Orion RISC processor;
can do fewer than 47000 pps; 8 Mbps
on one, maybe two, ports of a 4T card
but not on a 2T card.
[General note: Advanced 64 bit operating
systems (SGI IRIX, Digital Unix,
HP-Convex) have a usable memory limit of
roughly 16 GB ensuring that memory size
is not restricting router table growth.
Also note that adressable memory can be
much larger (SGI IRIX 1 TB) but can't be
used due to memory bandwidth restrictions;
sometimes also called memory latency;
a path requires about 110 bytes and a
BGP-4 route about 370 bytes of memory.
Generally the proccessing power limits
the number of routes since the required
power increases much faster than the
number of additional routes.]
Cisco 3620/3640 16 Mbps router with 2 or 4 expansion slots; 8
Mbps on one port of a 4T card.
Cisco 2650 37 kpps Cisco specs: 37,000 packets per second
Cisco 2620 25 kpps
Cisco 2610 15 kpps
Cisco 2500 series 3 Mbps entry level Cisco router; 3 Mbps
throughput 4000 pps if only one T1 port
is used; can handle 2 T1s if they are not
both fully loaded; 68030 CPU with up to
16 MB DRAM; BGP-4 possible, read Avi
Freedman's BGP tutorial; can forward
about 4000 packets per second (pps),
popular with ISPs.
DEC Alphastation 200 a pair is used as route servers at DEC's
NAP in Palo Alto with 128 MB DRAM each,
in 1997 a typical config was 27.9 MB
virtual memory size of gated with 6517 AS
paths and 34662 BGP-supplied routes.
DEC Gigaswitch 3.4 Gbps crosspoint switch, used at MFS and
MAE-West and other high volume exchange
points, tested under real world
conditions, limited by a backplane speed
of roughly 800 Mbps, aggregate throughput
3.6 Gbps; one of the few boxes available
with full-duplex FDDI interfaces.
Extreme Summit 7i 48 Mpps New record with 48 million IP packets per
second. It's a 64 Gbps backplane non-
blocking switch with 32 1 Gbps ports
which can all run at wire speed.
Supports routing protocols like BGP-4,
plus policy based Quality of Service
(QoS) bandwidth management and
priorization. The minimum latency is
about 5 microseconds.
Extreme Networks 12 Mpps It was a new record of 12 million packets
per second IP forwarding achieved with
the Summit 1 switch using hardware IP
routing "Hardware, hardware, sweet
hardware", 8 port Gigabit ethernet switch
with 17.5 Gbps backplane capacity and
switching latency under 6 microseconds,
better than ATM with OC12 622 Mbps
Foundry 178 Mpps Up to 120 Gigabit ports per chassis and
a total throughput of up to 480 Gbps and
178 million packets per second.
Ironbridge Networks 1.3 Tbps The SwiftCOR is the world's first Terabit
router and is scalable from gigabits up
to 1.3 terabits/sec. Supports up to 256
OC48s or up to 64 OC192 interfaces or any
combination thereof. Virtual IP backbone
Juniper M160 160 Gbps The M160 can forward up to 160 billion
bits per second using an ASIC and a
parallel forwarding and route computation
architecture. Mountain View, initial
investors: LM Ericsson, Northern Telecom,
Siemens, 3Com, UUnet/WorldCom, Lucent
Technologies, Newbridge Networks. IP
routing in silicon, not the traditional
CPU and software approach. Each of the
maximum of eight interface cards can do
up to 10 billion bits per second full
duplex. Packets can be priortized. Real
world 10 Gbps on an interface is the
current record among all boxes. SSH
available (FreeBSD based), "JUNOS"
Juniper M40 40 Mpps Juniper's "flagship" router is the
previous model M40 which can do 40 million
packets per second and is based on a
similar technology as the M160 described
Juniper M20 20 Mpps Internet backbone router purpose built
for emerging Internet service providers.
IPSILON invented IP switching, a method to
efficiently put TCP/IP over any network
fabric, routing calculation on AMD/Intel,
can be used e.g. over ATM, alliance with
DEC, developped IGRP, competitor to Cisco,
but seems to have flamed out recently and
was purchased by Nokia.
Livingston Enterprises 3 Mbps IRX routers have up to 2 T1/E1 sync ports
and 2 low speed 56/64k sync ports. RIPv1,
OSPF and recently BGP-4 became available.
Memory can be upgraded beyond 16 MB if
desired (but all four must be equal).
Livingston remote access servers are
popular with ISPs, for modem and ISDN
National 2.5 Gbps National introduces industry's first
LVDS-compliant 4X4 Crosspoint Switch:
DS90CP04. : The 2.5 Gbps speed allows
the new device to be used in a broad
variety of applications from OC12 (622
Mbps) to OC48 rates (STM-16)
NortelNetworks Accelar 15 Gbps Layer 3 switch, routing, filtering,
policy in silicon, Model 1200 has been
shipping for 12-18 mths, with 6 I/O slots,
dual-gig blades, (optionally linksafe)
SX or LX. Not completely non-blocking,
as 12 Gig ports at full duplex would
overrun the backplane of 15 Gbps.
New model 8600... 128 Gbps 8 slot or 4 slot chassis (plus a pair of
switch fabrics...) 8 port Gigabit cards
(SX, LX, XD) Current backplane = 64Gbps
per switch fabric, total therefore
128Gbps. Approx throughput 100 Mpps
NSC PS32 sometimes preferred by the military.
Proteon popular on backbones a couple years ago.
3Com 18 Mfps New Superstack II 9300 12 port gigabit
ethernet switch can forward up to 18
million frames per second.
Wellfleet used on MCI's DS3 based frame-relay
backbone, Wellfleet is now a part of Bay
Networks, and now Nortel Networks.
Gbps: Giga bits per second 1,000,000,000 bits per second
Mbps: Mega bits per second 1,000,000 bits per second
Mpps: Mega packets per second packets have about 200 bytes or 1600 bits
pps: packets per second but can vary from ~100 to ~1600 bytes
created in 1994
last update July 2002
Networking and Distributed Computing page