Conference Paper Reprint - Advanced Micro Devices

17386A1992 IEEE, Reprinted with permission, from Proceedings of IEEE Local Computer Networks, Minneapolis, Sept. 13–15, 1992.

Multimedia Over FDDI

Conference Paper Reprint

Advanced

Micro

Devices

Multimedia Over FDDI

Amit Shah1, Don Staddon2, Izhak Rubin3, Aleksandar Ratkovic4

Advanced Micro Devices1, IBM2, IRI Corp. and UCLA3, IRI Corp. and UCLA4

Abstract

This paper focuses on issues in multimedia networking

over FDDI and simulation results demonstrating the

feasibility of multimedia over FDDI. Previous work in this

field has concentrated on the asynchronous channel in

FDDI. Other examinations of the FDDI synchronous

channel have applied bursty traffic to the synchronous

channel. Multimedia traffic is typically stream oriented.

Our simulations were focused on simulating typical

configurations with a range of applications and traffic

models. The advantages of FDDI as an integrated

voice/video/ data network are demonstrated.

1: Introduction

The paper is organized as follows.

A brief discussion on the basics of FDDI and

multimedia is followed by an overview of the various

multimedia applications and issues in networking of

multimedia applications such as voice and video. The

importance of parameters such as network delay, intra-

packet jitter, and bandwidth is highlighted.

The simulation models are discussed. The methodology

involved testing the effect of variations in TTRT, network

size, packet size, and offered load on the multimedia

streams of voice and video, using the IRI Planyst‘ tool.

Finally, the simulation results for running multimedia

over the FDDI synchronous channel are discussed. A

comparison of multimedia over FDDI synchronous and

asynchronous channels is made demonstrating the ability of

the synchronous channel to provide a near-constant latency

and bandwidth under various loads and configurations.

1.1: FDDI Basics

FDDI is a 100 Mbps high speed local area network

standard developed under the auspices of American

National Standards Institute (ANSI) X3T9.5 committee.

Unlike other LANs whose origins were proprietary

products, FDDI was developed by a group of people whose

interest was to create a reliable fault-tolerant, high-speed

network connecting numerous stations over greater

distances than existing standards. The ANSI X3T9.5

committee thus developed a specification for a network

based on a dual counter-rotating fiber optic ring using

timed-token protocol, which is capable of transmitting data

at 100 Mbps in each ring and which can extend to 500

stations over a total fiber length of 200 km with full system

performance.

The FDDI MAC supports multiple classes of service:

Synchronous, Asynchronous and Restricted Asynchronous.

The asynchronous class is used for normal, bursty data

traffic. For more constant, isochronous traffic, the

synchronous channel can be used and it provides

guaranteed bandwidth to each station on each token

rotation. The restricted asynchronous service provides a

mechanism for extended and protected dialog between a

limited number of stations. This feature is rarely used.

There have been several publications on FDDI [10],

[11], [12], [13], [14], [15], [16] which explain the protocol

in further detail.

2: Multimedia

Multimedia is a term used to denote a set of

applications, products, and technologies [1]. We define it

as the use of multiple means to communicate information

via a computer. Whereas computers today are mainly used

for textual data information presentation and storage,

multimedia uses the computer for text, natural and animated

images, rendered graphics, auditory stimuli and realistic

sound. There are two key components to multimedia:

multiple output media and interaction.

Television as we know it today exemplifies one of the

components of multiple output media; it delivers 'realistic'

pictures, animation and sound. However, the content of

what is viewed, and the way it is presented is largely

beyond the control of the user. The user interaction is

limited to turning on or off, switching channels, volume

control and some image control.

Personal computers demonstrate the other key concept

in multimedia: interaction. Unfortunately, most current

personal computers exploit only a limited number of media

such as text, graphics and a few sound tones. It is only now

that the personal computers' ability to present clearer,

sharper images is being exploited by multimedia

applications.

2.1: Multimedia applications

The concept of multimedia is only as interesting as the

applications it can support. It is the scope of multimedia

applications which is drawing such a tremendous response

from the information technology industries.

The various multimedia applications are shown in table

1 below.

Table 1: Multimedia applications

APPLICATION MEDIA

A=Audio, V=Video,

T=Text, I=Images,

G=Graphics

DESCRIPTION

Education/ Training A, V, T, I A live or taped video instructional session locally at a station or

across a network. This involves transfer of audiovisual and

textual instructions.

Interactive Education/

Training A, V, T, I, G A live or taped video instructional session locally at a station or

across a network. This involves transfer of audiovisual and

textual instructions. The interaction may be via voice, text or

graphics.

Information kiosks A, V, T, G This is dispensing of information such as legal, tourist, consumer

catalogs, dictionaries etc... These kiosks typically have minimal

interaction with the user; mainly instructions and commands.

Banking T, I Number transfers along with documents such as checks, money

orders, internal papers etc...

Medical Info. systems V, I, T, A Stored or real-time capture of X-Rays, CAT scans, reports etc...

with ability to do history, compare and micro and macro

searches.

Library I, T, V Search and retrieval of texts, audio and video cassettes.

Real Estate A, V, I, T Ability to view a neighborhood, block, street, house and interior

of home remotely.

Electronic Mail T, A, I Ability to leave audiovisual messages along with textual

messages.

Home Video Distribution V, A Cable company maintains library of movies which can be

selected and played-back at viewer's convenience.

Travel Agents A, V, I, T Similar in concept to real estate with ability to book travel,

lodging and boarding in one call.

CAD/CAM T, G, A Standard engineering CAD / CAM with the ability to voice-

annotate and have on-line graphics help and notes as pop-up

boxes.

Electronic Collaboration A, V, T Video conferences, concurrent CAD/CAM etc...

Advertising V, A, I, T Ability to reconstruct, edit and create images annotated with

audio and text. The images may be sequenced into a video clip.

Weather I, T Ability to parse satellite images of the atmosphere into weather

reports.

There are several issues with multimedia, most notably

compression, synchronization and network transparency.

Multimedia applications typically consume a lot of

disk space for storage and current storage devices such as

Hard Disks, floppy disks, CD-ROMs are not capable of

storing massive amounts. Moreover, these storage devices

have a very slow transfer rate (read, write). Motion video

requiring 75 Mbps transfer rate cannot be sustained using

a CD-ROM without some compression techniques.

Synchronization of multimedia objects is an important

issue [8]. An object is a unit of data which may be a pixel,

encoded audio, the multimedia document itself, etc...

Synchronization can be at different granularity's. An

example of synchronization is voice and video streams;

the voice accompanying a video clip needs to be

synchronized to the picture. This issue is currently being

addressed in the multimedia community.

Most multimedia applications will need network

support [9]. Videoconferencing [7], distributed training,

etc... require the use of a network. Insofar as the network

is concerned, the only difference between multimedia and

other applications is the integration of voice, video and

data on the same network. Imaging can be modeled as

bursty data and hence does not require any additional

service of the network.

Almost all data transmission is asynchronous in nature.

It is unpredictable and of varying duration. Most local

area networks are optimized for high throughput, bursty

data transmissions over a shared medium with little or no

latency constraints. If the network is lightly loaded, and a

station applies a large load to the network, it will be able

to transmit the load with minimal delays. However, if the

network is heavily loaded, a station applying a large load

will encounter significant delays before transmission.

Thus, although Ethernet under light load offers excellent

average latencies, at high load it offers little or no bound

on the network access time.

Public networks or the telecom networks, on the other

hand, are typically optimized for circuit-switching

applications such as voice which requires low bandwidth

and low latency. These networks typically cannot provide

low access delay to bursty data.

In order to evaluate the feasibility of using existing

LANs in multimedia applications, we examine the issues

in sound and video transmissions such as bandwidth

requirements, latency, jitter and maximum number of

sessions. The characteristics of sound and video and the

requirements of multimedia on networks are examined.

Finally, the feasibility of multimedia over FDDI is

explored.

3: Definitions

Available bandwidth is the bandwidth which is actually

available for valid transmissions. Available bandwidth

can also be measured in terms of network efficiency.

Thus, in FDDI efficiency h = (T - D) / T; where T= target

token rotation time in ms and D = ring latency in ms.

If D=0.1 ms and T = 10 ms, then h = 99% and available

bandwidth is 99 Mbps whereas total bandwidth is 100

Mbps.

Latency is the average end to end message delay which

includes time for A/D conversion (if any), sample and

encode, packetization, queuing delay, transmission delay,

propagation delay, receive delay, decode, and

presentation.

Jitter is the maximum instantaneous variation in object

presentation time. If the object is a packet, then the

maximum inter-packet arrival time variation is defined as

packet jitter.

Session is defined as an interactive communications

dialogue between two or more users. Thus a telephone

conversation between two people is a session which

consumes a portion of the available bandwidth.

4: Characteristics of Sound

We classified sound as human speech and music.

Human speech or voice is typically in the 0-4 khz

spectrum. The bandwidth of music discernible by the

human ear is 20-25 khz (high fidelity systems have a

bandwidth of 22 khz).

Conversational sound (speech) consists of talk-spurts

followed by silence periods [3], [5], [6]. The ratio is

typically 35:65 respectively, with only one person

speaking at a time.

talk-spurt

350 ms silence

650 ms

Figure 1: Speech pattern

In digitized voice, the voice signal is sampled at the

Nyquist rate (twice the signal rate) in order to recover the

original signal correctly.

Therefore the voice sampling rate is 2*4 khz = 8 khz or

one sample every 125 ms in telephony. For stereo sound,

the audio frequencies extend up to 22 khz. Hence

sampling frequencies of up to 44 khz are also used in

digital stereo sound. Each sample is coded into a bit-

stream and the number of code-bits varies depending

upon the coding scheme used.

Table 2: Compressed digital audio streams

TRANSMISSION RATES bits/sample

2400 bps compressed 1-2

64 kbps standard PCM 6-10

40-16 kbps ADPCM scheme[22] 5-2

300-400 kbps stereo 8-10

Although telephone voice is not very bandwidth

intensive, stereo sound requires up to 0.4 Mbps

bandwidth. At lower bit rates, the audio-quality

deteriorates. Determining the quality of digitized audio is

a very subjective phenomenon and studies have indicated

a wide range of acceptable quality. It also depends on the

application. Spoken voice can be compressed

significantly before it becomes incomprehensible. The

study of stereo quality sound is even more subjective.

There are several coding and compression techniques

available for voice [2], [18], [21], [22]. Adaptive

Differential Pulse Code Modulation (ADPCM) and Digital

Speech Interpolation (DSI) are two of the popular

mechanisms for voice compression. DSI is well-suited for

packetized voice transmission as it conserves bandwidth

during silence intervals in a conversation.

4.1: Issues in sound transmissions

In voice transmissions [2],[3],[4],[5],[19], bandwidth is

not an issue in the LAN environment, although it may be

in stereo sound. A critical issue in interactive speech or

music (or voice mail) played back over the network is

latency. Once a person starts speaking, or music starts to

play, intermittent delays during the speech or music

session, are often noticeable, and sometimes annoying.

Therefore, once the session starts, it is desirable to

maintain a continuous stream of sound. In fact, as shown

in the table below, several studies have indicated that

maximum tolerable latencies for speech are of the order of

600 ms. Our experiences with satellite communications

has demonstrated that even 250 ms (one way) delays are

annoying though coherence is not impaired. For music,

the latencies may be more noticeable and hence the delays

may be required to be even less in order to be

imperceptible. Several studies have been conducted in

order to determine effects of network delays on voice

transmissions [3], [4], [5], [21], [31], [32].

Table 3: Effects of latency on human ear perception

One-way delay Effect of delay

>600 ms Speech becomes incoherent and unintelligible.

600 ms Speech is barely coherent.

250 ms Annoying. Conversation style has to be changed.

100 ms Imperceptible if listener hears from network only and not off the air.

50 ms Imperceptible even if the listener in same room and can hear naturally off the air and from

the network.

Using interactive speech as a model, we decided that

the maximum end-to-end tolerable latency was 100 ms.

These latencies would be acceptable for a large spectrum

of multimedia applications.

The only effect of token jitter is on the need to buffer.

In a truly isochronous network (providing constant

latency), there would be zero buffering requirement in the

network. In an asynchronous network, there is a need to

buffer. So long as the buffering is not excessive, the jitter

has minimal impact on system design. For example, if the

maximum packet latency and hence the maximum jitter

between packets were to be restricted to less than 15 ms

(as we targeted for the simulation), then

the maximum buffering required for a 64 kbps digital

voice stream would be 120 bytes, which is very small.

5: Characteristics of video

Video is moving pictures. It is different from imaging

and graphics mainly in the motion component. Video

represents motion scenes as a rapid sequence of still

frames. An NTSC compatible video is 640 x 480 at 30

frames per second. A PAL compatible video is 768 x 516

at 25 frames per second. The smaller the window size

(fewer number of lines scanned), the lower the bandwidth

requirement.

Table 4: Effect of frame rate on human eye perception

Frames per second

(fps) Effect on human eye

<10 fps Eye cannot discern motion. Each frame appears disjointed

12-15 fps Eye can discern motion but is jerky.

30 fps Television quality. Cannot discern high motion component

(blurred); e.g. baseball

60-75 fps HDTV quality. Can discern motion in high-motion games; e.g. ice-

hockey

90 fps Limit of human eye perception

1000 fps Scientific video quality; e.g. shuttle blast-off recording

The video signal can consume tremendous bandwidth.

An uncompressed digital NTSC video can consume

anywhere from 90 Mbps to 200 Mbps depending on the

encoding. This is enough to overrun any existing network

capacity. Compressed video [[28], [29], [30], [32] offers a

significant reduction in the bandwidth needs while

maintaining similar picture quality. Video bandwidth can

be reduced in several ways:

•Variable resolution

•Variable frame rate

•Reduced color fidelity

•Removing intra-frame and inter-frame redundancy

Table 5: Video compression techniques, rates and bandwidth requirements

Compression technique frame rate (fps) bandwidth

MPEG [30] 30 1.5 Mbps video stream

MPEG II 60-75 4-10 Mbps

P x 64 [28] 6-15 64 kbps - 2 Mbps

5.1: Issues in video transmissions

Digital video characteristics have not been studied as

well as digital voice. It is difficult to characterize

compressed video, video codecs and effect of network

delay. The issue of compression algorithms and

implementations is beyond the scope of this paper and the

parameter of most interest to us is the network delay.

Using a well-known maxim that " the human eye is

more forgiving than the human ear", we can apply the

restriction that video must meet the same delay constraints

as those of voice. Depending on the video application,

this delay may vary.

We selected 100 ms as an acceptable end-to-end

latency [31], [32]. We assumed that codecs at each end

will consume 30 ms in processing each frame and

outputting data (at MPEG rates). This leaves an effective

40 ms latency for the network component of the latency.

Of this, 10 ms is the latency across the WAN, and the rest

is the LAN component of the delay. Assuming

symmetrical LANs, this leaves 15 ms per LAN acceptable

latency.

WAN

FDDI FDDI

LAN latency

( 15 ms.)

End-to-end latency (100 ms.)

CODEC

latency

(30 ms)

CODEC

latency

(30 ms)

LAN latency

( 15 ms.)

WAN latency

(10 ms.)

Figure 2: Latency distribution across the network

If the maximum packet latency and hence the maximum

jitter between packets were to be restricted to less than 15

ms (as we targeted for the simulation), then the maximum

buffering required for a 1.5 Mbps stream would be

approximately 3 Kbytes, which is usual in an FDDI

adapter.

6: Use of FDDI Synchronous class of service

for video/voice applications

The purpose of our study was to examine the feasibility

of multimedia applications over the FDDI synchronous

channel when the FDDI is shared with normal, bursty data

traffic. The synchronous channel offers a protected, low-

latency bandwidth, which when unused is available to the

normal asynchronous transmissions.

A portion of the FDDI bandwidth is allocated to

synchronous services, either at startup or later by a

bandwidth allocater. Up to 100% of the network

bandwidth can be allocated to the synchronous service. In

other words, it is possible to implement a synchronous

only network. This allocation can be fixed, dynamically

allocated at session initiation, or on any granularity

preferred by the network administrator. A standard

allocator such as CCITT Q.931, may be used to perform

call-setup, tear-down and bandwidth allocation and

monitoring. SMT defines a protocol and several MIB

attributes which can be used to monitor and control the

bandwidth allocation [27].

Each multimedia station is allocated a portion of the

synchronous bandwidth. In order to allocate the

bandwidth, each station needs to characterize the

application requirements in terms of overhead and

payload, where overhead includes token capture, framing

and higher layer protocol headers, and payload is the

actual synchronous data (e.g. voice, video). This should

be calculated in units of bytes per 125 microseconds. An

application of 1.5 Mbps would require 1.5 x 106 x 125 x

10-6 /8= 23.4375 rounded up to 24 units of bandwidth. A

similar calculation should be done for the overhead. The

total bandwidth available is of the order of 100 x 106 x

125 x 10-6 /8 = 1562 units. Following the bandwidth

allocation, it is necessary to select the packet sizes for the

negotiated TTRT. For example, with a TTRT of 8 ms, the

packet size for the above synchronous traffic (1.5 Mbps

stream) can be calculated as the number of bytes that the

stream will produce in 8 ms. This is 1500 bytes. Hence,

with an 8 ms TTRT, and a 1.5 Mbps video stream, 1500

byte packets will be transmitted per token rotation.

The above method of allocating bandwidth was

selected so that an application would not have to change

bandwidth allocation every time that the TTRT value

changed. Only the packet size would change while

maintaining a constant overhead. There are other

mechanisms for allocating bandwidth which may be

simpler and more suitable for different network

configurations.

For the purposes of the simulation we allocated

synchronous bandwidth based on the TTRT value. Thus

an 8 ms TTRT with a 1.5 Mbps application would require

1500 byte transmission time or 1500 x 80 ns = 0.12 ms per

TTRT.

7: Network operation

An FDDI network can be operated in three ways:

1] asynchronous only;

2] mixed asynchronous and synchronous;

3] synchronous only.

We decided to test the network in all modes of

operation with a special emphasis on the mixed

asynchronous and synchronous mode. The offered

network load was varied from 80% to about 150% of

capacity. The traffic was a mixture of various applications

such as voice, video, imaging, file servers, and interactive

data.

We wanted to test the network not for its maximum

configurations but for its typical configurations. After

conducting a survey of the existing implementations and

practices, we concluded that a maximum unsegmented

network was in the range of 40-60 nodes. An

unsegmented network has nodes on the same physical

cable-plant with no intervening bridges, routers or some

such interworking units. Usually, networks do not exceed

50 nodes because of issues such as loading, administrative

domains, and traffic isolation. We selected a network

with nodes in the range of 48 to 55 as the representative

maximum of the typical network.

7.1 Topology

The following topology was adopted as the model for

the study. A LAN-WAN-LAN model was seen as

appropriate for typical multimedia services. We assumed

that the two LANs were symmetrical in behavior.

WAN

FDDI FDDI

concentrator/

station

Figure 3: Topology for simulation

A single hierarchy of connections was selected because

of the linearity of each LAN segment. Therefore, the

delay characteristics for a single LAN could be linearly

scaled to represent multiple-level LAN hierarchy. The

WAN connection was assumed to offer a fixed latency

path. The WAN could consist of ISDN, fractional T1, T1

or T3 lines depending on the bandwidth required.

7.2: Ring latency

A ring latency of 84 ms was used to test a small ring.

This corresponds to roughly 7 kilometers of cable. To test

the behavior of the network with larger ring latencies, a

ring latency of 1 ms was selected. This corresponds to

roughly 190 kilometers of cable.

7.3: Target Token Rotation Time

We decided to operate with three values of TTRT- 8

ms, 16 ms, and 24 ms TTRT values less than 8 ms were

not selected because network efficiency drops

significantly [33]. TTRT values larger than 24 ms were

not selected because the packet latencies would be

unacceptable for the multimedia traffic. Additionally, for

a synchronous only network, a TTRT of 26 ms was used.

7.4: Traffic and service models

Of the 50 stations, 26 were set-up to be voice/video

stations, 3 were low-rate imaging sources, 10 were

interactive data terminals, and 10 were file-servers.

Optionally, high burst-rate imaging sources (1 to 7

stations) were used in place of the 3 low-rate imaging

sources. The traffic was thus split into voice/video,

imaging and data. Imaging and data were further sub-

divided. There were two models for the imaging and two

models for the data.

The following table shows the characteristics of the

various traffic models that we selected.

Table 6: Traffic distributions

TRAFFIC TYPE INTER-

ARRIVAL

TIME (in ms)

PACKET

LENGTH

(in bytes)1

Peak OFFERED

LOAD

(in Mbps)2

Avg. OFFERED

LOAD

(in Mbps)

BUFFER

SIZE

(in packets)3

Imaging host 0.33 4K + 256 106.25 10 1000

Imaging workstation 3.6 4500 10 2 10

file data 37 1500 + 256 10 3 50

interactive terminal data 40 500 - 0.1 10

voice416, 84.5 64, 2028 + 20 + 256 0.032, 0.218 0.0112, 0.076 10, 10

video 8,14,16,24 1500, 2304, 3000, 4500 1.5, 1.316, 1.5, 1.5 1.5, 1.316, 1.5, 1.5 10

gateway5 - - 10 10 50

1 The length denotes the data + headers. The headers were deliberately chosen to be a large number.

2 The arrival rate for some traffic had a distribution model rather than a constant rate, which led to peak offered loads and average offered

loads.

3 The buffer size corresponds to the buffering at the transmit and receive queues. If an incoming packet finds the buffer full, it is dropped.

This corresponds to blocking.

4 An interactive voice model (32 Kbps ADPCM), and a MPEG stored voice-stream model were selected for modelling.

5 The gateway loading was approximately equal to 4 voice/video stations and 1.3 fileservers.

Overall, we stressed the network with a variety of traffic

models to ensure that the network is robust under extreme

operating conditions.

7.4.1: Video streams

Each video source represents a compressed video

stream. An MPEG-type stream rate of 1.5 Mbps is used.

This consists of a train of packets ranging from 4500 bytes

to 1500 bytes. The overall video loading is characterized

through a parameter which represents the number of

simultaneously active video sources.

A prescribed fraction of the video stream is directed

through a bridge or a gateway to another FDDI or WAN

network.

The target video latency for a video packet across a

single FDDI is set to be around 15 ms (99% of packets)

and of the order of 10 ms in the average.

7.4.2: Voice streams

Associated with each video stream is a packet voice

stream. This stream was characterized as interactive and

stored (MPEG CD-ROM specifications).

The interactive voice is characterized as a sequence of

32 kbps voice packets, each 64 bytes in size. The voice

source is modeled as a sequence of on-off periods,

representing the talk-spurts and silence periods typical in a

voice conversation. The ratio of talk to silence periods is

35: 65. The resulting offered load of the voice stream is

11.2 kbps.

The stored voice is characterized as a sequence of 2

Kbyte packets (2028 + 20 MPEG headers), with an inter

arrival rate of 84.5 ms. These packets are typically

interleaved with the video stream in a ratio of one voice

every six video packets. On the FDDI, these packets are

repacketized if necessary to smaller packet sizes.

The desired latency for voice packets across a single

FDDI is also set to be approximately 15 ms (99% of

packets) and of the order of 10 ms in the average.

7.4.3: Data : Interactive

This traffic consists of short packets (500 bytes)

arriving at random with low response time requirements of

50 ms (99% of packets) and 25 ms in the average.

7.4.4: Data -File transfer

This is modeled as file data to/from Ethernet hosts with

the FDDI being used as a backbone to Ethernet clients.

The file length is assumed to be uniformly distributed

between 1500 and 25000 bytes. Each file on average

consists of 8 packets of 1500 bytes + 256 bytes headers,

each arriving at 10 Mbps once the file transfer is initiated.

The file inter-arrival time is exponential. The offered

load is 3.6 Mbps.

7.4.5: Imaging : low-end (workstation)

This traffic source was used in some of the simulation

runs.

This is modeled as images coming off Ethernet hosts

into the FDDI host at 10 Mbps. The image size

distribution is uniform over 1.25 - 5 Mbytes. This image

stream is packetized into maximum length FDDI packets

(4096 + 256 bytes). The image inter-arrival time is varied

and a default value of 20% on-time and 80% off-time is

assumed. This assumes that a host is busy with imaging

only 20% of the time. Thus the peak offered load is 10

Mbps, but the average offered load is 2 Mbps.

The maximum acceptable delay in transmitting an

image and receiving it at the receiver is 1 s (99% of

packets) and 0.5 s average delay. A buffer size of 10

packets is used. Any over-flow leads to packet dropping.

7.4.6: Imaging : High-end (host)

This traffic source is used to simulate the impact of a

very bursty load on the network. The image distribution is

uniform over 1.25 to 5.625 Mbytes. A single image

stream consists of regularly arriving maximum sized

packets (4096 + 256 bytes). The average image inter

arrival rate is 0.32768 ms. The peak offered load is

106.25 Mbps and the average offered load is 10 Mbps.

8: Results

The following sections summarize the results of the

simulations.

8.1: Case 1- Asynchronous only network

In an asynchronous only network, with no synchronous

bandwidth allocated or used, the voice and video are

treated as data. No separate queue is allocated on transmit

or receive. In such a network also it is possible to

maintain a bound on the delay suffered by the packets.

The following observations refer to figures 4 to 19.

8.1.1: Effect on 99% latencies

Due to the unpredictable nature of the traffic

(asynchronous and bursty), the delay cannot be tightly

bounded. As can be seen from the figures 8 and 10, the

99% latencies suffered by video packets is as high as 48

ms when the network is not overloaded but running close

to capacity (90% load). When the network is overloaded,

the latencies can be as high as 252 ms.

In a more typical environment, where the traffic does

not consist of such high burst sources (imaging at 100

Mbps), it is possible to obtain low latencies. We were

able to verify this in our simulation (see figure 6) where in

an asynchronous only network, with 86% network

loading, and a TTRT of 24 ms voice and video packet

latencies were restricted to under 15 ms.

8.1.2: Effect of TTRT on latencies

In an overloaded network the higher the TTRT value,

lower the latencies. In the 8 to 24 ms range, it was

observed that the 24 ms TTRT value consistently offered

lower delays for all traffic types when the ring was large.

(figure 8 and 10).

8.1.3: Effect of ring size

Increasing ring latencies had an adverse effect on the

packet latencies. This was reflected in the increase in the

mean and maximum latencies for voice and video (figure

6 and 10). This effect was less noticeable in the 99%

latencies.

8.1.4: Effect of buffer sizes

Buffer sizes were allocated to the individual queues,

asynchronous and synchronous at different stations.

Therefore the asynchronous buffer size at the imaging

station was varied from 10 to 1000 packets, at the file

server stations it was 50 packets, and at the interactive

terminals it was 10 packets. Every synchronous station

had a 10 packet buffer.

These buffers never overflowed except in the overload

scenario. Then too, the blocking or buffer overflow was

occurring only at the imaging station. This result is

intuitive because the imaging stations were offering

instantaneous overload. This burst would fill up the

buffers and since the network was not faster than the

application, the buffers would not empty out fast enough.

At heavy loads these buffers are unable to empty and

hence lead to overflow. Low burst-rate applications such

as file service and interactive terminal traffic were not

capable of exceeding the network capacity and hence the

network could always clear the buffers faster than the

application could fill them.

The percentage of traffic blocked due to overflow

decreased slightly with increasing TTRT. This is again

intuitive because the larger TTRT values allowed longer

transmission times which in turn allowed the buffers to be

drained more often.

8.1.5: Effect on gateway

For traffic from FDDI to other LANs or WANs, the

latency was gated by the characteristics of the other LAN

or WAN. So long as the outbound traffic was less than the

capacity of the WAN or LAN, there was little or no

queuing delay. The major component of the delay was

then the transmission delay.

The effect of the changing of the various network

parameters on the gateway (figures 4-18) was similar to

that of other end-stations except that in the overload case

the gateway suffered significant blocking. The gateway

was the bottleneck for voice/video sessions spanning the

LAN-WAN-LAN connection. The minimum latencies

observed when the network loading was 90% and with at

least one high burst-rate source on the network, was 24 ms

at 8 ms TTRT.

When the high burst-rate source (imaging with peak

offered load of 100 Mbps and average of 10 Mbps) was

removed, the gateway provided acceptable performances

with latencies less than 15 ms.

8.2: Case 2- Asynchronous plus synchronous

network

In this network the voice and video are transmitted

over the synchronous channel. The synchronous

bandwidth is allocated per station based on the application

requirement. The voice and video packet sizes are varied

according to the TTRT requirements. If 64 byte

interactive voice packets are used, then the packet size is

constant for the different TTRT. For video, the packet-

size is 1500 bytes for 8 ms, 3000 bytes for 16 ms and 4500

bytes for 24 ms TTRT. The gateway is used for

voice/video and file server data forwarding. The gateway

is allocated synchronous bandwidth in proportion to the

traffic leaving the LAN. If four video streams are leaving

the LAN, then the gateway is allocated bandwidth equal to

four video streams (e.g. at 0.75 ms per video stream the

gateway is allocated 3 ms).

8.2.1: Effect on 99% latencies

We observe that the 99% latencies for voice/video

streams is fairly constant and under all circumstances-

90% and 150% loading on a small ring, 90% and 150%

loading on a large ring, and different TTRT - the 99%

latencies are within 24 ms. For 8 ms and 16 ms TTRT

values, the latencies are always within 16 ms. Even under

extreme stress, the synchronous channel offered a low-

latency path for time-critical applications such as

multimedia.

8.2.2: Effect of TTRT on latencies

Within the range of TTRT values yielding acceptable

latencies, it was more difficult to isolate the better TTRT.

We observed that while 8 ms TTRT yields excellent

values for voice/video traffic, the asynchronous bursty

traffic suffered lower latencies at the higher TTRT values.

Considering all traffic streams, we observed that a 16 ms

TTRT offered better all-round latencies in a mixed

synchronous and asynchronous network (figure 5, 7, 9 and

11).

8.2.3: Effect of ring sizes

To observe the effect of ring sizes on the network

performance, we simulated with a small ring size of 84 ms

latency and a large ring size of 1000 us latency. The

effect of increased ring sizes on voice / video (VV) is

readily apparent on the mean delays. The mean latencies

increased by as much as 40% whereas the 99% latencies

increased by about 10-15% only. Since the important

parameter for system design is the 99% latency rather than

the mean latency, this implies that in the range of typical

ring sizes (50 ms to 400 ms ), the 99% latencies are fairly

constant.

8.2.4: Effect of buffer sizes

The effect of buffer sizes is slightly higher blocking

than in the asynchronous only network. This is due to the

synchronous traffic effectively blocking some portion of

the bandwidth. If 20% of the bandwidth is allocated to the

synchronous channel, then for the asynchronous

applications FDDI appears to be a 80 Mbps asynchronous

data pipe. This is an excellent result because it implies

that if the asynchronous application capacity requirement

is known, the rest of the FDDI bandwidth can be allocated

to the synchronous channel and there will be little or no

effect on the asynchronous applications.

8.2.5: Effect on gateway

For traffic from FDDI to other LANs or WANs, the

latency was gated by the characteristics of the other LAN

or WAN. The major component of the delay was then the

transmission delay.

The effect of the changing of the various network

parameters was similar to that of other end-stations

(figures 5, 7, 9,11, 13, 15, 17). It was observed that for

the incoming traffic from the gateway, the latencies were

less than 16 ms for TTRT values of 8 and 16 ms, under

overload. For 24 ms TTRT the delay was of the order of

23 ms under overload. In light to heavy loading

conditions the latency was 8 ms.

We also observed that changing the number of

outbound sessions had little or no impact on the latency if

the appropriate synchronous bandwidth was allocated.

8.2.6: Synchronous bandwidth allocation

It was observed that incorrect bandwidth allocation

could dramatically affect the performance. Allocating the

appropriate bandwidth resulted in a guaranteed low-

latency channel for the application.

8.3: Synchronous only network

We simulated a synchronous only network with 70

voice/video stations and a ring latency of 122 ms (figure

19). In this configuration, the only traffic on the network

was voice/video. Each station was generating voice and

video traffic at a combined rate of 1.5 Mbps. Each station

was allocated 0.37 ms per token rotation leading to a

TTRT of 26 ms. The total load was near the maximum

sustainable by the network.

The results showed that the 99% latencies were less

than 7 milliseconds. In fact, even though the TTRT was

26 ms ( in order to accommodate 70 stations), the actual

maximum delay suffered by any packet was of the order

of 10 milliseconds.

It was also observed that there was a sharp drop in

performance with a small increase in the number of

stations. This is because the network is operating at

maximum capacity and even a slight increase in load can

cause the synchronous bandwidth to be over allocated.

We concluded that it is possible to calculate the maximum

number of synchronous stations that can be supported and

provide acceptable latencies. For the given application

(1.5 Mbps MPEG stream), it is between 55-65 stations,

depending on the safety factor desired.

It is not advisable to apply bursty and asynchronous

traffic to the synchronous channel as it can lead to

extremely high delays [34], [35]. This is because a burst

offers instantaneous load which can exceed the allocated

bandwidth causing the queuing component of the delay to

increase significantly .

9: Conclusion

The FDDI asynchronous mode provides excellent

multimedia services at normal loads. However, it has

limitations under heavy and extremely bursty loads which

may not be acceptable for voice/video services. The FDDI

synchronous mode of transmission provides a near

constant, low latency service under various loads and

configurations. Our results demonstrate that a large

number of simultaneous MPEG and Px64 sessions can be

supported even when the network is in over-load . Audio,

video, imaging and data communications services can be

integrated over FDDI without compromising any service

requirement in all but the most extreme cases. These

results demonstrate the feasibility of FDDI as an

integrated network and are part of a continuing study on

multimedia over FDDI.

10: Acknowledgments

Our acknowledgments to Bob Grow (XDI), J. D. Russell

(IBM), Bob O'Hara, Basil Alwan and Dave Roberts

(Advanced Micro Devices) for their invaluable help and

suggestions in the model development.

11: Bibliography

[1] I. Brackenbury & H. Sachar, "The road to

multimedia," Future Generation Computer Systems 7,

1991, pp. 91-96

[2]K. Sriram, R. S. McKinney & M. H. Sherif, "Voice

Packetization and Compression in Broadband ATM

Networks," IEEE JSAC, vol. 9, no. 3, Apr. 91, pp. 294-

303

[3]T. Bially, A. J. McLau ghlin, & C. J. Weinstein, "Voice

communication in integrated digital voice and data

networks," IEEE Trans. Commun., vol. COM-28, Sep. 80,

pp. 1478-1490

[4]J. G. Gruber, "Delay related issues in integrated voice

and data networks," IEEE Trans. Commun., June 81, pp.

787-800

[5]J. G. Gruber, "Performance Requirements for

integrated Voice/Data networks," IEEE JSAC, DEC. 83,

pp. 981-1005

[6]P. T. Brady, "A statistical analysis of on-off patterns in

16 conversations," Bell Syst. Tech. J., vol. 47, Jan. '68

[7]K. Watabe, S. Sakata,, et. al, "Distributed Desktop

Conferencing System with Multiuser Multimedia

Interface," IEEE JSAC, vol. 9, no. 4, may 91, pp. 531-539

[8]R. Steinmetz, "Synchronization Properties in

Multimedia Systems," IEEE JSAC, vol. 8, no. 3, Apr. 90,

pp. 401-412

[9]C. Nicolaou, "An Architecture for Real-Time

Multimedia Communication Systems," IEEE JSAC, vol.

8, no. 3, Apr. 90, pp. 391-400

[10]D. Dykeman & W. Bux, "Analysis and Tuning of the

FDDI Media Access Control Protocol," IEEE JSAC, vol.

6, no. 6, July 88

[11]R. M. Grow, "A timed token rotation protocol for

local area networks," Electro '82, Token Access Protocols,

May 82

[12]M. J. Johnson, "Reliability mechanisms of the FDDI

high bandwidth token ring protocol," Comput. Networks

ISDN Syst., voi. 11, Feb. 86, pp. 121-131

[13]M. J. Johnson, "Proof that timing requirements of the

FDDI token ring protocol are satisfied," IEEE Trans.

Commun., vol. COM-35, June 87, pp. 620-625

[14]F. E. Ross, "FDDI - a Tutorial," IEEE

Communications, May '86, 24(5)

[15]V. Iyer & S. Joshi, "FDDI's 100 Mbps Protocol

Improves on 802.5 Spec's 4 Mbps LImit," EDN, May 2

'85, pp. 151-160

[16]R. K. Moulton, "FDDI - a 100 Mps Solution on

Fiber," Lightwave, Oct '86, pp. 47-54

[17]H. R. Miller, P. Zafiropulo & F. Closs, "Data/Voice

integration based on IEEE 802.5 Token-Ring LAN,

EFOC/LAN 86 Proceedings, June '86, pp. 66-76

[18]N. S. Jayant & S. W. Christensen, "Effects of Packet

Losses in Waveform Coded Speech," IEEE Trans. on

Communications, COM-299(2), Feb '81

[19]M. Frontini & G. Watson, "An investigation of

packetised voice on the FDDI token ring", HP technical

report, no. HPL-BRC-TR-87-019

[20]V. S. Frost, W. W. LaRue, et. al, "A tool for local area

network modeling and analysis," SIMULATION, nov. '89,

pp. 283-297.

[21]D. O. Bowker, C. B. Armitage, "Performance issues

for packetized voice communications," Proc. Nat.

Commun. Forum, vol. 41, n0. 3, '87, pp. 1087-1092

[22]M. H. Sherif, D. O. Bowker, et. al., "Overview of

CCITT embedded ADPCM Algorithms," Proc. IEEE

Supercomm/ICC'90, Atlanta, GA, Apr. '90, vol. 3, pp.

1014-1018

[23]T. D. C. Little & A. Ghafoor, "Network

Considerations for Distributed Multimedia Objects

Composition and Communication," IEEE Network

Magazine, Nov. '90, pp. 32-49

[24]FDDI Token Ring Media Access Control," American

National Standard for Information Systems, X3.139-1987

[25]FDDI Physical Layer Protocol, American National

Standard for Information Systems, X3.148-1988

[26]FDDI Physical Layer, Medium Dependent (PMD),

American National Standard for Information Systems,

X3.166-1990

[27]FDDI Station Management (SMT), Preliminary Draft

proposed American National Standard, Rev. 7.1,

X3T9/92-037

[28]CCITT Recommendation H.261, Line Transmission

on Non-telephone Signals: Video Codec for Audiovisual

services at p x 64 kbps

[29]ISO/IEC DIS 10918-1; Information technology -

Digital compression and coding of continuous-tone still

images; a.k.a JPEG

[30]ISO/IEC JTC 1/SC 29 N 071; Coded representation of

picture, audio and multimedia/hypermedia information;

a.k.a MPEG

[31]J. D. Russell, "Trends in Communication Throughput

Demand," Fibertour/ComputerNet Conf., Boston, 10/90

[32]J. D. Russell, "Communications Performance

Requirements in Multimedia Applications"

[33]R. Jain, "Performance Analysis of FDDI Token Ring

Networks: Effect of Parameters and Guidelines for setting

TTRT," IEEE Lightwave Telecommunications Systems,

May '91, pp. 16-22

[34] W. Genter & K. S. Vastola, "Delay Analysis of the

FDDI Synchronous Data Class", IEEE

[35]P. Martini & T. Meuser, "Service Integration in

FDDI," Proceedings of the 15th Conference on Local

Computer Networks, 1990, IEEE Computer Society Press

Mean,99% Delay (20,21,22- Image WS Not Shown)

Figure 4: Async. only n/w, normal load, 84 microsec. ring

Voice Video Interactive File Data Video Gtw

Mean

99%

TTRT (ms)

OfLd .9, Async, 16.8km.

Mean,99% Delay (23,24,25)

Figure 5: Sync./Async. n/w, normal load, 84 microsec. ring with imaging ws

100

1000

10000

Voice Video Imaging WS Interactive File Data Video Gtw

Mean

99%

TTRT (ms)

TTRT 8- SBA: .25(VV), 1.0(G

TTRT 16- SBA: .5(VV), 2.0(G

TTRT 24- SBA: .75(VV), 3.0(G

OfLd .9, Async/Sync, 16.8km.

Mean,99% Delay (20,21,22)

Figure 6: Async. only n/w, normal load, 84 microsec. ring with Imaging

100

1000

10000

Voice Video Imaging

WS Interactive File Data Video Gtw

Mean

99%

TTRT (ms)

OfLd .9, Async, 16.8km.

Figure 7: Sync./Async. n/w, normal load, 84 microsec. ring

Voice Video Interactive File Data Video Gtw

Mean

99%

TTRT (ms)

TTRT 8- SBA: .25(VV), 1.0(GW

TTRT 16- SBA: .5(VV), 2.0(GW

TTRT 24- SBA: .75(VV), 3.0(G

OfLd .9, Async/Sync., 16.8km.

Figure 8: Async only n/w, normal load, 1000 microsec. ring

100

110

120

130

Voice Video Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

OfLd .9, Asyn, 200km.

Mean, 99% Delay (35,36,37- Image WS not shown)

Figure 9: Sync/Async n/w, normal load, 1000 microsec. ring

100

110

120

Voice Video Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

TTRT 8- SBA: .22(VV), .88(GW)

TTRT 16- SBA: .488(VV), 1.95(GW

)

TTRT 24- SBA: .75(VV), 3.0(GW)

OfLd .9, Async/Sync, 200km

Figure 10: Async only n/w, normal load, 1000 microsec. ring with Imaging WS

100

1000

10000

Voice Video Imaging

WS Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

32 38.4

3584

2048

51.2

86.4

128

OfLd .9, Asyn, 200km.

Figure 11: Sync + Async n/w, normal load, 1000 microsec. ring with Imaging WS

100

1000

10000

Voice Video Imaging WS Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

TTRT 8- SBA: .22(VV), .88(GW)

TTRT 16- SBA: .488(VV), 1.95(GW)

TTRT 24- SBA: .75(VV), 3.0(GW)

14.4

22.4 22.4

3840

2150

1792

115

22.4

14.4

OfLd .9, Async/Sync, 200km.

Figure 12: Async. only n/w, overload, 84 microsec. ring

100

110

120

130

140

150

160

170

180

190

200

210

220

230

240

250

260

270

280

290

300

310

320

330

340

350

360

370

Voice Video Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

4%bl

13%bl

16%bl

60%bl

55%bl

60%bl

OfLd 1.5, Async, 16.8km.

Figure 13: Sync + Async n/w, overload, 84 microsec. ring

100

110

120

130

140

150

160

170

180

190

200

210

220

230

240

250

260

270

280

290

300

310

320

Voice Video Interactive File Data Video Gtw

TTRT ms

Mean

99%

TTRT 8- SBA: .25(VV), 1.0(G

TTRT 16- SBA: .5(VV), 2.0(G

TTRT 24- SBA: .75(VV), 3.0(G

8%bl

7%bl

15%bl

OfLd 1.5, Async/Sync, 16.8km.

Figure 14: Async. only n/w, overload, 84 microsec. ring with Imaging WS

100

1000

10000

Voice Video Imaging WS Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

4%bl

13%bl

16%bl

60%bl

55%bl

60%bl

50%bl

45%bl

OfLd 1.5, Async, 16.8km.

Figure 15: Sync. + Async n/w, overload, 84 microsec. ring

100

1000

10000

100000

Voice Video Imaging WS Interactive File Data Video Gtw

TTRT ms

Mean

99%

TTRT 8- SBA: .25(VV), 1.0(G

TTRT 16- SBA: .5(VV), 2.0(G

TTRT 24- SBA: .75(VV), 3.0(

55%bl

60%bl

55%bl

8%bl

7%bl

15%bl

OfLd 1.5, Async/Sync, 16.8km.

Figure 16: Async only n/w, overload, 1000 microsec. ring with Imaging WS

100

1000

10000

100000

Voice Video Imaging WS Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

61%bl

65%bl

50%bl

14%bl

12%bl

19%bl 66%bl

62%bl

OfLd 1.5, Async, 200km.

Figure 17: Sync + Async n/w, overload, 1000 microsec. ring with Imaging WS

100

1000

10000

100000

Voice Video Imaging WS Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

65%bl

64%bl

58%bl

25%bl

12%bl

18%bl

OfLd 1.5, Async/Sync, 200km.

Figure 18: Async only n/w, normal load, 1000 microsec. ring

100

110

120

130

140

150

160

170

180

190

200

210

220

230

240

250

260

270

280

290

300

310

320

330

340

350

360

370

380

390

400

410

420

430

Voice Video Interactive File Data Video Gtw

TTRT (ms)

Mean

99%

14%bl

12%bl

19%bl

66%bl

62%bl

OfLd 1.5, Async, 200km.

Figure 19: Sync only n/w, normal load, 122 microsec ring

Voice Video

Mean

99%

All VV, All Synch

70 Stations Suppo

Utilizing 96.8 Mbp

TTRT: 26 ms

SBA: 0.37 ms

OffLd .96, Sync Tfc, 24km.