3.3 Cell Switching (ATM)

Asynchronous Transfer Mode (ATM) connection-oriented packet-switched network used in both WAN and LAN settings Q.2931: signaling (connection setup) protocol

discover a suitable route across an ATM network allocate resources at the switches along the

circuit (to ensure the circuit a particular quality of service)

Cells packets are called cells (fixed length)

53-byte cell = 5-byte header + 48-byte payload commonly transmitted over SONET

other physical layers possible

Variable vs. fixed-length packets no optimal packet length

if small: high header-to-data overhead if large: low utilization for small messages

variable-length packets lower bound

minimum amount of information that needs to be contained in the packet

typically a header with no optional extensions

upper bound (set by a variety of factors) the maximum FDDI packet size, for example,

determines how long each station is allowed to transmit without passing on the token, and thus determines how long a station might have to wait for the token to reach it

fixed-length packets it is easier to build hardware to do simple jobs, and the

job of processing packets is simpler when you already know how long each one will be

if all packets are the same length, then each switching element takes the same time to do its job

packet vs. cell most packet-switching technologies use variable-length

packets cells are fixed in length and small in size

Big vs. Small Packets

Small improves queue behavior cell: finer control over the behavior of queues examples

cell cell length = 53 bytes link speed = 100Mbps the longest wait ≈ 53 × 8/100 = 4.24μs [exactly

4.04μs] for ATM

variable-length packets maximum packet length = 4KB link speed = 100Mbps transmission time ≈ 4096 × 8/100 = 327.68μs

[exactly 312.5μs] a high priority packet that arrives just after the

switch starts to transmit a 4KB packet will have to sit in the queue 327.68μs waiting for access to the link

cell: shorter queue length than that of packets when a packet begins to arrive in an empty queue,

the switch have to wait for the whole packet to arrive before it can start transmitting the packs on an outgoing link

it means that the link sits idle while the packet arrives

if you imagine a large packet being replaced by a “train” of small cells, then as soon as the first cell in the train has entered the queue, the switch can transmit it

example variable-length packets

if two 4-KB packets arrived in a queue at about the same time

the link would sit idle for 327.68μs while these two packets arrive

at the end of that period we would have 8KB in the queue

Cell if those same two packets were sent as trains of

cells, then transmission of the cells could start 4.24μs after the first train started to arrive

at the end of 327.68μs, the link would have been active for a little over 323μs (≈ 327.68μs – 4.24μs)

there would be just over 4 KB of data left in the queue, not 8KB as before

shorter queues mean less delay for all the traffic

Small improves latency (for voice) voice digitally encoded at 64Kbps (8-bit samples at 8KHz) need full cell’s worth of voice samples before transmitting a

cell a sampling rate of 8KHz means that 1 byte is sampled every

125μs (= (1/8000) * 103), so the time it takes to fill an n-byte

cell with samples is n × 125μs a 1000-byte cells implies 125ms to collect a full cell of

samples before you even start to transmit it to the receiver

Cell Format

Two different cell formats User-Network Interface (UNI) format

host-to-switch format interface between a telephone company and one of its

customers

Network-Network Interface (NNI) format switch-to-switch format interface between a pair of telephone companies

GFC HEC (CRC-8)

4 16 3 18

VPI VCI CLPType Payload

384 (48 bytes)8

ATM cell format at the UNI

User-Network Interface (UNI) GFC (4 bits): Generic Flow Control(not wildly used)

Provide a means to arbitrate access to the link if the local site used some shared medium to connect to ATM

VPI (8 bits): Virtual Path Identifier VCI (16 bits): Virtual Circuit Identifier

For now, we can think of them as a single 24-bite identifier that is used to identify a virtual conncection

Type (3 bits): management, user data CLP (1 bit): Cell Loss Priority

A user or network element may set this bit to indicate cells that should be dropped preferentially in the event of overload

User-Network Interface (UNI) ….

HEC (8 bits): Header Error Check (CRC-8) Protecting the cell header is particularly important because an error

in the VCI will cause the cell to be misdelivered

Network-Network Interface (NNI) GFC becomes part of VPI field (no GFC and becomes 12-bit

VPI)

ATM Headers

Architecture of an ATM network

Segmentation and Reassembly Segmentation and Reassembly (SAR)

in ATM, the packets handed down from above are often larger than 48 bytes, and thus, will not fit in the payload of an ATM cell

solution fragment the high-level message into low-level

packets at the source transmit the individual low-level packets over

the network reassemble the fragments back together at the

destination

Segmentation is not unique to ATM But it is much more of a problem than in a

network with a maximum packet size of, say, 1,500 bytes

To address the issue, ATM Adaptation Layer was added

ATM Adaptation Layer (AAL) a protocol layer sits between ATM and the variable-

length packet protocols that might use ATM, such as IP

the AAL header simply contains the information needed by the destination to reassemble the individual cells back into the origins message

■ ■ ■ ■ ■ ■

AAL

ATM

AAL

ATM

Segmentation and reassembly in ATM

Because ATM was designed to support all sorts of services, including voice, video, and data, it was felt that different services would have different AAL needs

ATM Adaptation Layer (AAL) AAL1 and 2 designed for applications that need

guaranteed bit rate (e.g., voice, video) AAL 3/4 designed for packet data

AAL3 used by connection-oriented packet services (such as X25)

AAL4 used by connectionless services (such as IP) AAL5 is an alternative standard for packet data

AAL3 and AAL4 are merged into one known as AAL ¾, so there are now four AALs

ATM Layers

ATM Layers in Endpoint Devices and Switches

AAL 3/4

AAL3/4 provide enough information to allow variable-length packets

to be transported across ATM network as a series of fixed-length cells

AAL supports the segmentation and reassembly process the task of segmentation / reassembly involves two different

packet formats Convergence Sublayer Protocol Data Unit (CS-PDU)

[AAL3/4 packet format] ATM cell format for AAL3/4

Convergence Sublayer Protocol Data Unit (CS-PDU) PDU (Protocol Data Unit), a new name for packet CS-PDU defines a way of encapsulating variable-length

PDUs prior to segmenting them into cells the PDU passed down to the AAL layer is encapsulated by

adding a header and a trailer [CS-PDU header and a trailer], and the resultant CS-PDU is segmented into ATM cells

CPI Btag BASize Pad 0 Etag Len

8 16 0─24 8 8 16< 64 KB8

User data

PDUAAL 3/4 packet

(CS-PDU)ATM cell

encapsulate(header &

trailer)

segment

Encapsulation and segmentation

CS-PDU format CPI (8 bits): Common Part Indicator (version field)

Only the value 0 is currently defined

Btag/Etag (8 bits): Beginning and Ending tag BASize (Buffer Allocation) (8 bits): a hint to the reassembly

process as to how much buffer space to allocate for the reassembly

Pad: make sure the data is one byte less than a multiple of 4 bytes

0-filled byte ensures the trailer is 32 bite Len (16 bits): length of the PDU

CPI Btag BASize Pad 0 Etag Len

8 16 0─24 8 8 16< 64 KB8

User data

CPI Btag BASize Pad 0 Etag Len

8 16 0─24 8 8 16< 64 KB8

User data

ATM Adaptation Layer 3/4 packet format

In addition to CS-PDU header and trailer, AAL 3/4 specifies a AAL 3/4 header and trailer [AAL 3/4 header and trailer]

The CS-PDU is segmented into 44-byte chunks, an AAL 3/4 header and trailer is attracted to each one, bringing it up to 48 bytes, which is then carried as the payload of an ATM cell

CS-PDUheader

CS-PDUtrailer

User data

44 bytes 44 bytes 44 bytes 44 bytes

ATM header

AAL header

Cell payload

AAL trailer

Padding

ATM Cell Format (AAL3/4) Type (2 bits) SEQ (4 bits)

sequence number detect cell loss or misordering

MID (10 bits) multiplexing identifier multiplex several PDUs onto a single connection

Payload (352 bits;44 bytes) the segmented CS-PDU chunk

ATM header Length CRC-10

40 2 4

SEQ MIDType Payload

352 (44 bytes)10 6 10

AAL3/4 Type field

Length (6 bits) number of bytes of PDU contained in this cell it must be 44 for BOM and COM cells

CRC-10 error detection anywhere in the 48-byte cell payload

ATM header Length CRC-10

40 2 4

SEQ MIDType Payload

352 (44 bytes)10 6 10

ATM header Length CRC-10

40 2 4

SEQ MIDType Payload

352 (44 bytes)10 6 10

ATM cell format for AAL3/4

5 bytes

48 bytes5 bytes

2 bytes 44 bytes 2 bytes

ATM header

Payload

AAL 3/4 trailer44-byte chunkAAL 3/4 header

ATM header

Encapsulation and segmentation for AAL3/4 the user data is encapsulated with the CS-PDU header

and trailer the CS-PDU is then segmented into 44-bye payloads,

which are encapsulated as ATM cells by adding the AAL3/4 header and trailer as well as the 5-byte ATM header

the last cell is only partially filled whenever the CS-PDU is not an exact multiple of 44 bytes

CS-PDUheader

CS-PDUtrailer

User data

44 bytes 44 bytes 44 bytes 44 bytes

ATM header

AAL header

Cell payload

AAL trailer

Padding

CS-PDUheader

CS-PDUtrailer

User data

44 bytes 44 bytes 44 bytes 44 bytes

ATM header

AAL header

Cell payload

AAL trailer

Padding

Encapsulation and segmentation for AAL3/4

With 44 bytes of data to 9 bytes of header, the best possible bandwidth utilization would be 83%

AAL5

PDUAAL 5 packet

(CS-PDU)ATM cellencapsulate segment

CS-PDU Format Data portion trailer (8-byte)

2-byte Reserved + 2-byte Len + 4-byte CRC-32

Pad (up to 47 bytes) so trailer always falls at end of ATM cell

Length size of PDU (data only)

CRC-32 (detects missing or misordered cells)

CRC-32

< 64 KB 0─47 bytes 16 16

ReservedPad Len

32

Data

CRC-32

< 64 KB 0─47 bytes 16 16

ReservedPad Len

32

Data

ATM Adaptation Layer 5 packet format

Encapsulation and segmentation for AAL5 user data is encapsulated to form a CS-PDU the resulting PDU is then cut up into 48-byte

chunks, which are carried directly inside the payload of ATM cells without any further encapsulation

User data

48 bytes 48 bytes 48 bytes

ATM header Cell payload

Padding

CS-PDUtrailer

User data

48 bytes 48 bytes 48 bytes

ATM header Cell payload

Padding

CS-PDUtrailer

Encapsulation and segmentation for AAL5

3.3.3 Virtual Path ATM uses a 24-bit identifier for vircuit circuits

8-bit virtual path identifier (VPI) 16-bit virtual circuit identifier (VCI)

Example a corporation has two sites that connect to a public ATM

network, and that at each site the corporation has a network of ATM switches

we could establish a virtual path between two sites using only the VPI field

within the corporate sites, however, the full 24-bit space is used for switching

Public network

Network BNetwork A

Public network

Network BNetwork A

Example of a virtual path

Advantage of virtual path although there may be thousands or millions of

virtual connections across the public network, the switches in the public network behave as if there is only one connection

there needs to be much less connection-state information stored in the switches, avoiding the need for big, expensive tables of per-VCI information

TP 、 VPs 、 and VCs

Example of VPs and VCs

Connection Identifiers

Virtual Connection Identifiers in UNIs and NNIs

ATM Cell

Routing with a Switch

3.3.4 Physical Layers for ATM

In the ATM standard, it is assumed that ATM would run on top of a SONET physical layer many ATM-over-SONET products

Actually, these two are entirely separable, e.g., lease a SONET link from a phone company and

send whatever you want over it, including variable-length packets

send ATM cells over many other physical layers instead of SONET, e.g., over Digital Subscriber Line (DSL) links of various types

3.4 Implementation and Performance

A very simple way to build a switch buy a general-purpose workstation and equip it with

a number of network interfaces run suitable software to receive packets on one of

its interfaces perform any of the switching functions send packets out another of its interfaces

I/O bus

Interface 1

Interface 2

Interface 3

CPU

Main memory

I/O bus

Interface 1

Interface 2

Interface 3

CPU

Main memory

A workstation used as packet switch

The figure shows a workstation with three network interfaces used as a switch a path that a packet might take from the time it

arrives on interface 1 until it is output on interface 2

I/O bus

Interface 1

Interface 2

Interface 3

CPU

Main memory

we assume DMA (Direct Memory Access) the workstation has a mechanism to move data directly

from an interface to its main memory, i.e., direct memory access (DMA)

once the packet is in memory, the CPU examines its header to determine on which interface the packet should be out it then uses DMA to move the packet out to the

appropriate interface the packet does not go to the CPU because the CPU

inspects only the header of the packet

I/O bus

Interface 1

Interface 2

Interface 3

CPU

Main memory

Main problem with using a workstation as a switch its performance is limited by the fact that all packets

must pass through a single point of contention in the example shown, each packet crosses the I/O bus

twice and is written to and read from main memory once

the upper bound on aggregate throughput of such a device is, thus, either half the main memory bandwidth or half the I/O bus bandwidth, whichever is less (usually it’s the I/O bus bandwidth)

I/O bus

Interface 1

Interface 2

Interface 3

CPU

Main memory

example a workstation with a 133-MHZ, 64-bit wide I/O bus can

transmit data at a peak rate of a little over 8 Gbps (= 133 × 220 × 64)

since forwarding a packet involves crossing the bus twice, the actual limit is 4 Gbps

this upper bound also assumes that moving data is the only problem a fair approximation for long packets a bad one when packets are short

the cost of processing each packet- (1) parsing its header and (2) deciding which output link to transmit it on-is likely to dominate

example, a workstation can perform all the necessary processing to switch 1 million packets each second (packet per second (pps) rate)

if the average packet is short, say, 64 bytes throughput = pps × (bits per packet)

= 1 × 106 × 64 × 8 (bits per second)

= 512 × 106 (bits per second) this 512 Mbps would be shared by all users connected to

the switch example, a 10-port switch with this aggregate throughput

would only be able to cope with an average data rate of 51.2 Mbps on each port

To address this problem a large array of switch designs that reduce the

amount of contention and provide high aggregate throughput

some contention is unavoidable if every input has data to send to a single output, then

they cannot all send it at once if data destined for different outputs is arriving at

different inputs, a well-designed switch will be able to move data from inputs to outputs in parallel, thus increasing the aggregate throughput

Switchfabric

Controlprocessor

Outputport

Inputport

3.4.1 Ports

Switch

fabric

Control

processor

Output

port

Input

port

A 4 × 4 switch

The 4 × 4 switch in the figure consists of ports (input ports and output ports)

communicate with the outside world contain fiber-optic receivers and buffers to hold packets that

are waiting to be switched or transmitted, and often a significant amount of other circuitry that enables the switch to function

switch fabric when presented with a packet, deliver it to the right output

port control processor (at least one)

in charge of the whole switch

Switchfabric

Controlprocessor

Outputport

Inputport

Input port the first place to look for performance bottlenecks has to receive a steady stream of packets, analyze

information in the header of each one to determine which output port (or ports) the packet must be sent and pass the packet on to the fabric

Another key function of ports: buffering it can happen in either the input or the output port it can also happen within the fabric (sometimes

called internal buffering)

simple input buffering has some serious limitations example, an input buffer implemented as a FIFO as packets arrive at the switch, they are placed in the

input buffer the switch then tries to forward the packets at the front

of each FIFO to their appropriate output port if the packets at the front of several different input

ports are destined for the same output port at the same time, then only one of them can be forwarded; the rest must stay in their input buffers

Switch

2

21

Port 1

Port 2

Simple illustration of head-of-line blocking

drawback (head-of-line blocking) occurs at input buffering those packets left at the front of the input buffer prevent

other packets further back in the buffer from getting a chance to go to their chosen outputs

buffering wherever contention is possible

input port (contend for fabric) internal (contend for output port) output port (contend for link)

Switch

2

21

Port 1

Port 2

3.4.2 Fabrics

Should be able to move packets from input ports to output ports with minimal delay and in a way that meets the throughput goals of the switch Means fabrics display some degree of parallelism

Parallelism a high-performance fabric with n ports can often

move one packet from each of its n ports to one of the output ports at the same time

Types of fabric shared bus shared memory crossbar self-routing

Shared bus found in a conventional workstation used as a switch the bus bandwidth determines the throughput of the switch,

high-performance switches usually have specially designed busses rather than the standard busses found in PCs

Shared memory packets are written into a memory location by an input port

and then read from memory by the output ports the memory bandwidth determines switch throughput, so

wide and fast memory is typically used in this sort of design it usually uses a specially designed, high-speed memory bus

Crossbar a matrix of pathways that can be configured to

connect any input port to any output port in their simplest form, they require each output port

to be able to accept packets from all inputs at once Main problem: each port would have a memory

bandwidth equal to the total switch throughput

A 4 × 4 crossbar switches

Self-routing rely on some information in the packet header to direct

each packet to its correct output usually a special “self-routing header” is appended to the

packet by the input port after it has determined which output the packets needs to go to

this extra header is removed before the packet leaves the switch

self-routing fabrics are often built from large numbers of very simple 2×2 switching elements interconnected in regular patterns (i.e., banyan switching fabric)

001

011

110

111

001

011

110

111

000001010011

100101110111

Switch fabric

Output port

Input port

Original packetheader

Switch fabric

Output port

Input port

Self-routing header

Switch fabric

Output port

Input port

(a)

(b)

(c)

A self-routing header is applied to a packet at input to enable the fabric to send the packet to the correct output, where it is removed: (a) packet arrives at input port; (b) input port attaches self-routing header to direct packet to correct output (c) self-routing header is removed at output port before

packet leaves switch

Banyan Network constructed from simple 2 x 2 switching elements self-routing header attached to each packet elements arranged to route based on this header

look at 1 bit in each self-routing header route packets toward the upper output if it is zero or

toward the lower output if it is one

001

011

110

111

001

011

110

111

000

001

010

011

100

101

110

111

if two packets arrive at the same time and both have the bit set to the same value, then they want to be routed to the same output and a collision will occur

the banyan network routes all packets to the correct output without collisions if the packets are presented in ascending order

Routing packets through a banyan network. The 3-bit numbers represent values in the self-routing headers of four arriving packets.

001

011

110

111

001

011

110

111

000

001

010

011

100

101

110

111