Experiments with packet switching of voice traffic P.N. Clarke, B.Sc, Ph.D., and Prof. L.F. Turner, B.Sc, Ph.D. Indexing terms: Telephone exchanges and networks, Voice traffic Abstract: There has been much interest recently in integrated services digital networks carrying both voice and data traffic. Packet switching is being used to carry data in an attempt to make better use of trunk capacity than with circuit switching. In a telephone conversation, for most of the time only one person is talking, and it has been suggested that packet switching can lead to economies in carrying voice traffic also. In view of the variable delays associated with store and forward switching, buffering is usually required at the receiver to enable received speech to be reconstituted at the proper rate. Simulation experiments of packet switching of voice traffic with fixed packet routing have been carried out. The results of these simulation experiments, which are described in this paper, show that, for a single link between two exchanges, 22 conversations can be carried by packet switching with reasonable delay. For the same inter-exchange-link capacity, only 15 conversations can be carried by circuit switching. For a larger network with more exchanges and links per path, a similar advantage is also found with packet switching. The results show that the standard deviation of interpacket delay for successive packets of the same talkspurt is an order of magnitude less than the standard deviation of packet transit time for all packets. This suggests correlation of flows of packets within the same talkspurt. The wider variation of transit delay applies to each talkspurt as a whole and all packets within the talkspurt have correlated transit times, and hence interarrival times. The fact that the standard deviation of interpacket delay is small as compared with the standard deviation of packet transit time suggests that the receiver buffering requirement is less than that indicated by the standard deviation of the packet transit time. 1 Introduction As a result of the recent increases in data traffic, various suggestions have been put forward relating to the use of separate data networks. The existing analogue circuit- switched telephone network has transmission and noise characteristics which vary significantly through the network, and call set-up times of the order of seconds are involved. Although this situation is acceptable in so far as voice traffic is concerned, it is unacceptable for many data applications. On account of the burst-like nature of the data, in many applications store and forward switching methods, such as packet switching, have been proposed and implemented [1-6]. Packet switching makes better use of expensive high-capacity interexchange trunks by trans- mitting small blocks of data, or packets, only when there are data to be sent. If there are, for short periods, more packets for transmission than can be dealt with, some are stored for forwarding in later less busy periods. Packet switching makes efficient use of trunk capacity at the expense of variable delay. Rather than have two separate networks, one for data and one for voice traffic, a single network for both types of traffic may be more economical. As digital transmission and switching methods are being used increasingly for speech, and as data are best handled in digital form, inte- grated services digital networks (ISDN) are being pro- posed. These might be implemented using the new electronic digital circuit-switching exchanges, such as in System X [7]. Alternatively, depending on the relative costs of switching and transmission, packet switching might be used to make use of trunk capacity during silent periods. Systems such as TASI [8] have been used in the past on both transocean cable and satellite circuits in order to make use of silent periods. Packet switching with its variable delays might be con- sidered unsuitable for real-time application such as conver- sational speech. If, however, a buffer is used at the receiver Paper 2586G, first received 3rd June 1982 and in revised form 17th February 1983 Professor Turner is, and Dr. Clarice was formerly, with the Department of Electrical Engineering, Imperial College of Science & Technology, South Kensington, London SW7 2BT, England. Dr. Clarke is now with British Telecommunications, Gower Street, London WC1 E6BA, England then the variations in packet arrival times can be smoothed out and the received speech reconstituted at the correct rate. This does, of course, add to the total speech delay. The total delay resulting from packet creation, network transit time, and receiver buffering and decoding must not be too long (cf. 270 ms 1-way delay through a satellite link). It has been observed [9] that delay in excess of 900 ms can give rise to considerable difficulties. Replies and nonverbal responses, together with their relative timings, provide the speaker with clues as to the listener's understanding and thus aid the conversation process. Minoli [10, 11] considered theoretically talker behav- iour and end-to-end, that is, packet transit delay for a link packet-switched voice system. He also considered delay dependencies on packet size and the effects of the number of queue buffers at the link output. Coviello [12] also considered end-to-end delay for a variety of network par- ameters and a variety of alternative network protocols to facilitate packet switching of voice traffic. Gruber [13] reviews a variety of switching techniques for voice traffic and is again concerned with end-to-end delays. A variety of speech coding techniques are reviewed and the results of some ARPA network voice experiments are described by Gold [14]. These works [10-14] have been concerned very largely with the end-to-end, or packet transit, delay and its varia- tion, and the workers involved have considered this varia- tion to be the principal factor determining the buffering requirement at the receiver; with the buffer being necessary to even out irregular packet arrivals. Although the packet transit time, if large, and its variation may have a signifi- cant effect on conversational behaviour (see Reference 9), it is, however, the variation of interpacket delay, rather than packet transit time, which determines the receiver buffering requirement. In the experiments carried out and described in this paper the interpacket delay (that is, the delay between arrivals of successive packets within the same talkspurt) and its standard deviation were measured in order to investigate the correlation of packet flows. The results of simulation experiments carried out with a fixed packet routing system show that the standard deviation of interpacket delay for successive packets of the same talk- spurt is an order of magnitude less than the standard devi- IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983 105
Transcript
Experiments with packet switching ofvoice traffic
P.N. Clarke, B.Sc, Ph.D., and Prof. L.F. Turner, B.Sc, Ph.D.
Indexing terms: Telephone exchanges and networks, Voice traffic
Abstract: There has been much interest recently in integrated services digital networks carrying both voice anddata traffic. Packet switching is being used to carry data in an attempt to make better use of trunk capacitythan with circuit switching. In a telephone conversation, for most of the time only one person is talking, and ithas been suggested that packet switching can lead to economies in carrying voice traffic also. In view of thevariable delays associated with store and forward switching, buffering is usually required at the receiver toenable received speech to be reconstituted at the proper rate. Simulation experiments of packet switching ofvoice traffic with fixed packet routing have been carried out. The results of these simulation experiments, whichare described in this paper, show that, for a single link between two exchanges, 22 conversations can be carriedby packet switching with reasonable delay. For the same inter-exchange-link capacity, only 15 conversationscan be carried by circuit switching. For a larger network with more exchanges and links per path, a similaradvantage is also found with packet switching. The results show that the standard deviation of interpacketdelay for successive packets of the same talkspurt is an order of magnitude less than the standard deviation ofpacket transit time for all packets. This suggests correlation of flows of packets within the same talkspurt. Thewider variation of transit delay applies to each talkspurt as a whole and all packets within the talkspurt havecorrelated transit times, and hence interarrival times. The fact that the standard deviation of interpacket delay issmall as compared with the standard deviation of packet transit time suggests that the receiver bufferingrequirement is less than that indicated by the standard deviation of the packet transit time.
1 Introduction
As a result of the recent increases in data traffic, varioussuggestions have been put forward relating to the use ofseparate data networks. The existing analogue circuit-switched telephone network has transmission and noisecharacteristics which vary significantly through thenetwork, and call set-up times of the order of seconds areinvolved. Although this situation is acceptable in so far asvoice traffic is concerned, it is unacceptable for many dataapplications. On account of the burst-like nature of thedata, in many applications store and forward switchingmethods, such as packet switching, have been proposedand implemented [1-6]. Packet switching makes better useof expensive high-capacity interexchange trunks by trans-mitting small blocks of data, or packets, only when thereare data to be sent. If there are, for short periods, morepackets for transmission than can be dealt with, some arestored for forwarding in later less busy periods. Packetswitching makes efficient use of trunk capacity at theexpense of variable delay.
Rather than have two separate networks, one for dataand one for voice traffic, a single network for both types oftraffic may be more economical. As digital transmissionand switching methods are being used increasingly forspeech, and as data are best handled in digital form, inte-grated services digital networks (ISDN) are being pro-posed. These might be implemented using the newelectronic digital circuit-switching exchanges, such as inSystem X [7]. Alternatively, depending on the relativecosts of switching and transmission, packet switchingmight be used to make use of trunk capacity during silentperiods. Systems such as TASI [8] have been used in thepast on both transocean cable and satellite circuits inorder to make use of silent periods.
Packet switching with its variable delays might be con-sidered unsuitable for real-time application such as conver-sational speech. If, however, a buffer is used at the receiver
Paper 2586G, first received 3rd June 1982 and in revised form 17th February 1983Professor Turner is, and Dr. Clarice was formerly, with the Department of ElectricalEngineering, Imperial College of Science & Technology, South Kensington, LondonSW7 2BT, England. Dr. Clarke is now with British Telecommunications, GowerStreet, London WC1 E6BA, England
then the variations in packet arrival times can besmoothed out and the received speech reconstituted at thecorrect rate. This does, of course, add to the total speechdelay. The total delay resulting from packet creation,network transit time, and receiver buffering and decodingmust not be too long (cf. 270 ms 1-way delay through asatellite link). It has been observed [9] that delay in excessof 900 ms can give rise to considerable difficulties. Repliesand nonverbal responses, together with their relativetimings, provide the speaker with clues as to the listener'sunderstanding and thus aid the conversation process.
Minoli [10, 11] considered theoretically talker behav-iour and end-to-end, that is, packet transit delay for a linkpacket-switched voice system. He also considered delaydependencies on packet size and the effects of the numberof queue buffers at the link output. Coviello [12] alsoconsidered end-to-end delay for a variety of network par-ameters and a variety of alternative network protocols tofacilitate packet switching of voice traffic. Gruber [13]reviews a variety of switching techniques for voice trafficand is again concerned with end-to-end delays. A varietyof speech coding techniques are reviewed and the results ofsome ARPA network voice experiments are described byGold [14].
These works [10-14] have been concerned very largelywith the end-to-end, or packet transit, delay and its varia-tion, and the workers involved have considered this varia-tion to be the principal factor determining the bufferingrequirement at the receiver; with the buffer being necessaryto even out irregular packet arrivals. Although the packettransit time, if large, and its variation may have a signifi-cant effect on conversational behaviour (see Reference 9), itis, however, the variation of interpacket delay, rather thanpacket transit time, which determines the receiver bufferingrequirement. In the experiments carried out and describedin this paper the interpacket delay (that is, the delaybetween arrivals of successive packets within the sametalkspurt) and its standard deviation were measured inorder to investigate the correlation of packet flows. Theresults of simulation experiments carried out with a fixedpacket routing system show that the standard deviation ofinterpacket delay for successive packets of the same talk-spurt is an order of magnitude less than the standard devi-
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983 105
ation of the packet transit time. This thus suggests that thereceiver buffering requirements are significantly less thansuggested by packet transit-time statistics.
This paper describes an investigation into the delaysinvolved in the use of packet switching for voice traffic. Inthe course of the investigation, a computer simulationmodel was devised and this is described in Section 2 of thepaper. The experiments carried out and the resultsobtained are described in Section 3, and some conclusionsto be drawn from the work are presented in Section 4.
2 Packet-switched voice network simulationmodel
The simulation model developed will be described in twoparts:
(a) the talker activity model (Section 2.1)(b) the packet-switched network model (Section 2.2).
Part (a) deals with the nature of the interaction betweenthe talkers, and part (b) with the packet-switched networkitself, which transports the speech in packet form.
2.1 Talker activity modelIn most conversational speech between two people, one issilent at any given time (listening while the other istalking). There are, however, occasions when both aresilent and or when both are talking simultaneously (e.g.when one person interrupts the other). Talker activity canbe thought of in terms of active periods (talking) or silenceperiods. These periods can be the main active periods ofsignificant utterances, such as sentences, and the silence ofa listener while another person is talking. Alternatively, thefine structure of the significant utterances can be takeninto account. This fine structure refers to the actual timeduring which a sound is being made by a talker and thepauses between sentences, words and syllables.
The principal object of a packet-switched network is tomake efficient use of network transmission capacity. It isthus clear that packets should only be carried by thenetwork for any conversation, while either of the parties ofthat conversation is actually speaking. In this way, thesilence periods of conversation can be filled in on the high-capacity trunks which are shared by many talkers. A largernumber of talkers can thus use a given trunk capacity thanwith circuit switching. Speech detection equipment shouldproduce an output to be put into packets according to thecoarse or the fine structure of talker activity, depending onthe speech-detector sensitivity and switching speed.
Studies have been carried out of the talker activityduring telephone calls. Norwine and Murphy [15] con-sider principally the coarse structure of the interactionsbetween talkers. Brady describes an experimental arrange-ment for measuring fine structure of talker activity [16],the analysis of data gathered using this apparatus [17] andthe fitting of such data to a theoretical model for gener-ating probabilities of transition between states of talking,silence, interruption etc. [18].
The talker activity model used for the simulationexperiments, and reported on in this paper, was based onthe results given in Figs. 3 and 5 of the paper by Norwineand Murphy [15], and thus does not take account ofpauses within talkspurts. It would have been possible,using a Markov chain model, to obtain finer details oftalkspurt activity, but as this approach is considerablymore difficult to implement than the probability densityfunction approach, it was not adopted in the simulationsleading to the results presented in this paper. However, anapproach involving the consideration of the finer details of
talkspurt activity may well be of value, and could form thebasis of a more extensive further consideration of packet-switched systems used for the transmission of voice traffic.In Reference 15, graphs are given of talkspurt length andresponse time distributions, with response time beingdefined to be the length of time between the end of onetalker's talkspurt and the beginning of the next talker'stalkspurt. The distribution of response time includes nega-tive values, that is, interruptions. A positive value ofresponse time corresponds to the more normal period ofmutual silence between talkspurts before the next talkerbegins. Using the talkspurt duration statistics given in Ref-erence 15, the talkspurt duration was approximated in thework reported on in this paper, using a lognormal dis-tribution [19] having the same mean and modal values.
The lognormal distribution has a PDF,/(x), given by
/(*) =1
exp 2a
where /i and a2 are the parameters of the distribution.With the mean and mode of the distribution, as given inReference 15, n = 0.485 and a2 = 1.871. As regardsresponse time, this was approximated using a normal dis-tribution with mean 0.32 and standard deviation 0.584 (alltimes in seconds).
With the model used, the talker activity, which isdefined to be:
In the model used in the simulation, a talker was allowedto talk for a talkspurt length, with the length being drawnfrom the lognormal distribution. The response time for thesecond talker was drawn from the normal distribution.After the talkspurt length, the first talker stops, and thesecond talker is allowed to begin at a time equal to thesum of the talkspurt length and the response time after thestart of the first talker's talkspurt. The length of the talk-spurt for the second talker was determined from the log-normal distribution. In this way the times for the secondtalker to stop and for the first talker to begin again weredetermined. If as a result of a combination of interruptionsand long talkspurts a talker was scheduled to start a newtalkspurt during the course of an existing talkspurt, it wasarranged for the current talkspurt to be completed beforethe start of the next, which was then allowed to beginimmediately afterwards. These points are illustrated by thesimple example shown in Fig. 1.
In the Figure; at time A, talker 1 (Tl) begins to speakuntil B. T2 is idle at time A and is scheduled (by Tl) tostart speaking at C. At time B, Tl stops and becomes idleand T2 is idle but waiting to start at time C. At C, T2begins to speak until E and schedules Tl, who is idle, tostart at time D. This represents an interruption by Tl whowill start talking before T2 has finished. At time D, Tlbegins to speak until H. T2 is scheduled by Tl to start hisnext talkspurt at time F which is thus an interruption ofTl. T2 stops talking at E and awaits a new start at F. Attime F, T2 interrupts Tl and schedules Tl's talkspurt tostart at /. T2 stops talking at G and Tl carries on until H.Tl stops at time H and remains idle until the next start at/. At time /, Tl begins to speak until time M and schedulesT2 to start at time J. T2 starts speaking at J, interruptingTl and schedules Tl to start at time L. T2 stops at time K.
106 IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
talker 1(T1)
talker 2(T2)
talkspurtlength
responsetime I(•ve) |
|response. time(-ve)
I II II I
I!
I I '
X X
time
G H
Fig. 1 Example of talker-activity model
At L, Tl is still talking, so he continues the current talk-spurt (until M) and restarts immediately until N. Also attime M, T2 is scheduled to start at time 0, and so on.
In the simulation, the following procudure was adopted.During talkspurts, the speech from talker's equipment wastaken as having been digitised with all talker pairs in thenetwork having the same speech bit rates. When enough8-bit (byte) speech digits to fill a packet had been receivedfrom a talker, a packet was created at the exchange. Anappropriate header was added to the packet which thenwent for transmission through the network. The nextpacket of the talkspurt was then filled up, and so on. Atthe end of the talkspurt, the packet which was being filledup was completed by filling with 'blank' information at thespeech bit rate (see Fig. 2). All speech packets in thenetwork were thus of the same length. All packets as wellas being of the same length were created at regular inter-vals during the talkspurt.
Clearly, this simple model of the coarse structure oftalker activity, and regular packet generation, makes noallowance for the possibility, depending on the nature ofthe interruption, of a talker stopping when interrupted. Noallowance was made in the simulation model for the effectson talker behaviour of delay in packet creation, of cross-network delays, nor of buffering and speech reconstructiondelays. All of these delays will in general be variable,except the regular packet creation delay. Delays in tele-phone channels do affect talker behaviour, as has beenreported by Brady [20] in the case of fixed delays. Thesimple model was chosen to provide approximate conver-sational talker activity.
talkeractivity
time
packet-generationtimes
1 3
Fig. 2 Talker activity and packet creation
maximum data content of packet (bits)
Speech bit rate (bit/s)
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
J K L M N 0
key
talking
The rationale behind the simplfied approach was that ifthis model which does not allow for delays in speech, andoperates by generating full packets at regular intervals, canhandle more calls than a circuit switched system of thesame trunk capacity, then a more complicated model,allowing for delays and pauses within talkspurts, mayallow even more calls to take place.
2.2 Packet-switched network modelThe network of the simulation model was made up ofpacket-switching exchanges (PSEs) connected by full-duplex trunks. The talkers were connected to theexchanges by lines which can be assumed to be either ana-logue or digital (operating at the speech bit rate). In all theexamples, each talker was associated with another talker atanother PSE in the network. These talker pairs wereassumed to be engaged in conversation before the start ofeach experiment.
On generation of a speech packet at a talker's interface,the required outgoing trunk was determined by consultingthe route table. Fixed routing was used in all of the experi-ments. Packets entering the network from a talker couldonly be put into the queue for this trunk if there were morethan two free queue buffers. This gives some priority totransit traffic, i.e. to packets which have been accepted intothe network, for example, at node 4 in Fig. 4b, or at node 2in Fig. Ac for packets between 1 and 3, and between 3 and1. If an originating packet could not be accepted, it washeld in a buffer associated with the talker's interface to thenetwork. That talker's identity was put into a queueassociated with the trunk output queue. Whenever apacket was sent along the trunk and a queue bufferbecame free, the list of talkers with waiting packets wasinspected. If there were sufficient buffers to allow in anoriginating packet, the first one waiting joined the trunkoutput queue. If that talker had further waiting packets, herejoined the list of talkers with waiting packets.
Packets were transmitted over the trunks at the trunkbit rate. Copies of all transmitted packets were kept,pending acknowledgments received from the other end ofthe trunk. Associated with each copy of a packet, kept inthe retransmission queue, was a time by which that packetmust be acknowledged. This time was based on the worstpossible case of acknowledgment delay. Acknowledgedpackets were deleted from the retransmission queue. If a
107
packet were to exceed its time in the retransmission queue,then it would be retransmitted, followed by its successors(unless these had meanwhile been deleted) before any newpackets were transmitted. However, as transmission errorswere not simulated, the only condition under which theretransmission procedure could have been evoked was thatin which a packet was discarded at a transit node becauseof there being no free buffers in the output queue.
The acknowledgment process was carried out using thesend-and-receive sequence numbers carried by all packetsas used in the ISO's HDLC and in the CCITT's X.25recommendation [21, 22]. Any packet carrying a sendsequence number greater than that expected was discardedand a REJ (Reject) packet sent in the reverse direction.This REJ packet indicated the last correctly receivedpacket and instructed retransmission to start at the appro-priate point in the packet sequence. Only one REJ wasallowed in a given direction until the next expected packetwas received. If a REJ was corrupted by noise and thusdiscarded, the correct packet sequence was maintained byretransmission invoked by the timeout mechanism. In thecase of no outgoing packets when one was correctlyreceived, a RR (receiver ready) packet indicating correctreception was sent. This reduced the use of the timeoutmechanism under conditions of light trunk loading.
Packets made their way through the network to theirdestination. Here they were assumed to be passed to thereceiver interface for conversion to speech (after any buf-fering, if necessary). On arrival of every packet, the packetstatistics were updated. Packet statistics measuredincluded:
(i) the number of packets received(ii) the mean and standard deviation of packet transit
time for the packets of (i). Packet transit time was mea-sured as the difference between the arrival time at thedestination PSE and the packet creation time at the sourcePSE
(iii) the mean and standard deviation of packet inter-arrival time. Packet interarrival time was defined as thedifference between arrival times of successive packets of thesame talkspurt.
The simulation program was written in Simula [23, 24]and was designed to be as flexible as possible. A widevariety of networks and conditions could be simulated bychoosing appropriate input data for the program. Theinput data required for this were:
(i) the number of PSEs(ii) the number of trunks(iii) for each trunk: (a) the source and destination of
PSEs, {b) the trunk capacity, (c) the bit error probability,and (d) the retransmission timeout period
(iv) the route table (this gives the next PSE en route toeach destination)
(v) the talkers' speech bit rate (the same for all talkers)(vi) the speech packet length (in bytes)(vii) the number of PSE pairs with conversations
between them(viii) for each of (vii) above, the number of talker pairs(ix) the duration of the simulation and intervals between
statistics report(x) the seed for the random-number stream used.
3 Packet-switched voice network experiments
3.1 General descriptionThree simulation experiments were carried out, and therewere several model parameters common to the experi-ments. The maximum trunk output queue length was ten
packets (with two reserved for transit traffic). The talkerspeech bit rate* was 9600 bit/s. There were no local calls(i.e. calls between talkers at the same PSE). The programwas run for 250 s in all experiments and the results of thefirst 100 s were removed in order to reduce the bias effectsof no packets being present in the network at the start ofthe simulation. Analysis of the results has shown that astable condition was reached in this time. Results were alsocollected for a single talker pair. The three experimentscarried out were as follows:
(i) Two PSEs, one 144 kbit/s trunk (see Fig. 3); 128 + 8(speech + header) byte packets; 25 ms retransmissiontimeout interval; varying number of talker pairs.
(ii) Two PSEs; one 144 kbit/s trunk (see Fig. 3); 15, 20and 25 talker pairs, packet sizes of 32, 48, 64, 96, 128 (fromprevious experiment) 192 and 256 bytes (with 8 bytes ofheader in addition with correspondingly adjustedretransmission time.
(iii) Three PSEs (see Fig. 4): (a) a fully connectednetwork with 144 kbit/s trunks; (b). a star network with288 kbit/s trunks; and (c) a linear network with 288 kbits/strunks; 128 + 8 byte packets; 12.5 ms retransmission
PSE1trunk
PSE 2to
talkers
Fig. 3 2-node packet-switched voice network
1 2 3
Fig. 4 3-node packet-switched networks
a Three nodes: FC, b Three nodes: star, c Three nodes: linear
* A 9600 bits/s speech rate was used in order to facilitate the simulation. Thesignificance of the results so obtained is not, however, restricted by this. Appropri-ate time scaling of packet lengths and trunk-line rates would render them applicableat a more realistic speech data rate of 64 kbit/s.
108 IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
timeout period on the faster 288 kbits/s trunks of (b) and(c); number of talker pairs varied. (The abbreviation FC isused in the Figures to refer to the fully connected networkconfiguration.)
3.2 Results of experimentsThe results of the experiments will now be described.Related points in all Figures are joined by straight-linesegments to identify related points in the multigraphFigures and to indicate trends, rather than to show exactbehaviour, between the experimental points.
3.2.1 Two PSEs, 128 + 8 byte packets, varying numberof talker pairs: The number of packets transferred in the150 s (for each value of talker load) of the experiment forall talker pairs and for the single talker pair are shown inFig. 5. The number of packets for all pairs rises almostlinearly up to 25 talker pairs, with a smaller rise between25 and 27. For the single talker pair, almost the samenumber of packets are carried at all loads (total number oftalker pairs). The average packet transit time of Fig. 6shows little increase up to 22 talker pairs but shows anincreasing rate of increase above 22 pairs. The averagepacket transit time must be added to the packet creationtime of (128 x 8/9600) s = 107 ms to obtain the total delaybetween speech being uttered and becoming available forreconstruction on arrival at the destination PSE. Any buf-fering to allow for variations in arrival times must beadded as well. Up to 22 talker pairs, the transit time is lessthan 20 ms. The standard deviation of packet transit time,shown in Fig. 7, is low (less then 30 ms, suggesting receiverbuffering of over 100 ms) up to 22 talker points, but itincreases more rapidly as more talkers are added to thenetwork. This suggests that up to 22 talker pairs withspeech bit rate of 9600 bit/s, with 128 + 8 byte packets,can share a 144 kbit/s trunk with an average speech delay
40
'g35
^ 3 0
£ 25
In
I 2 0a> 1 5
I 1 05
talker pairsall
10
Fig. 5length =
0.9
0.8
a," 0.7
Z 06
I 0 5
Z 0.4
| 0.3
& 0.2
| 0.1° 0
•single15 20
talker pairs25 30
Packets transferred, varying talker load, two nodes, packet128 + 8
talker pairsall
single
10 15 20 25talker pairs
Fig. 6 Packet transit time, varying talker load
1EE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
30
of 107 + 20 = 127 ms (before receiver buffering). For morethan 22 talker pairs, the packet transit time and standarddeviation will lead to even greater delays. It will be noticedthat there is a difference between the packet transit timeand standard deviation curves for all talker pairs and forsingle talker pair (see Figs. 6 and 7). This is because of theeffects of the smaller sample size of packets from the singletalker pairs (see Fig. 5). The single talker pair results willnot be considered in the rest of this paper. Packet switch-ing appears to be able to carry the conversations of 22talker pairs (under the above conditions) before delaysbecome unacceptable. A 144 kbit/s trunk operating undercircuit switched conditions can carry (144000/9600)= 15conversations with no variable delay.
2.0in
1.? 1.'5
o
a
0.5
talker pairsall
single
10 15 2520
talker pairs
Fig. 7 SD of packet transit time, varying talker loads
30
The average interpacket delays of Fig. 8 are dominatedby the 107 ms packet creation delay and are almost equalto it for up to 25 talker pairs. In fact, the variation in delaydue to queueing was found to be approximately threeorders of magnitude less than the transit time, and toexhibit no systematic variations.! This indicates that, evenwith the extra transit time (queueing for transmission overthe trunk), there is little difference between the admissionqueueing and transmission delays for successive packets oftalkspurts. The increase in interpacket delay for more than25 talker pairs indicates that successive packets of eachtalkspurt take longer to reach their destination than eachof their predecessors. A packet-switched network is clearlyunsuitable for carrying speech traffic when operated in thisregion. The standard deviation of interpacket delay is lessthan 4 ms for less than 22 talker pairs. This is approx-imately an order of magnitude less than the standard devi-ation of packet transit time.
0.15ro.u-0.13-
(/> 0.12
gr o n3 0.10Z 0.09SL 0.08a 0.07£ 0.06c 0.05'Z 0.04 -§T0.03% 0.02o 0.01 -
talker pairsall
•^single
10 15 20talker pairs
25 30
Fig. 8 Interpacket delay, varying talker loads
t Details of the effects of speech statistics on the perception of impairments arisingfrom variable delays can be found in References 13 and 25.
109
This suggests greater correlation between arrivals ofsuccessive packets of the same talkspurt than indicated bythe transit-time figures. The packets are generated atregular intervals and, for long talkspurts (with respect topacket creation time), the packets of all active talkers arecorrelated. This correlation is disturbed slightly when atalker stops or when another joins the set of active talkers.Once the transit delay is determined, all the packets of thesame talkspurt have similar transit times, and thus inter-packet delays are similar. The transit time indicates thedelay in admission queueing and transmission, and thusrepresents the storage of packets within the network.
Variation in this for each talkspurt does not affect theinterarrival times of the packets and hence the speechreconstruction. The standard deviation of the packettransit time is thus a measure of the spread of time spent inthe network. The standard deviation of interpacket delay isthe measure which should be used in deciding on receiverbuffering requirements. Several times this standard devi-ation should be allowed in the receiver buffer to minimisethe number of late packets which will have to be dealt within some way.
The trunk utilisation, shown in Fig. 10, rises approx-imately linearly with the number of talker pairs. For lessthan 22 talker pairs, the utilisation is less than 80%. Formore than 22, the utilisation is greater than this, with theassociated rapid rise in transit times, as can be seen fromthe packet transit time of Fig. 6. It should be noted thatcertain parameters such as the number of packets trans-mitted and the trunk utilisation (which are shown in Figs.5 and 10, for example) can be calculated from a knowledgeof talker activity, packet generation rate, packet length,trunk capacity and the number of talkers. It should also benoted that the number of talker pairs required to achieve100% trunk utilisation can be calculated using the packetgeneration rate, talker activity and trunk capacity. If this is
done, then it is found that 31 talker pairs are required, andthis agrees with extrapolation of Fig. 6. However, it is clearthat before this number of talker pairs is actually reachedthe variations in delays are such as to render speech trans-mission unacceptable.
3.2.2 Two PSEs, varying packet sizes: The differentpacket sizes used in this experiment, with their creationtimes and retransmission timeout periods on the 144 kbit/strunk, are shown in Table 1. Totals of 15, 20 and 25 talkerpairs were used in the experiment with each value ofpacket length (except for 25 pairs 32 + 8 bytes).
Table 1 : Packet details
Packet length Packet creationtime
Retransmissiontimeout
bytes32 + 8 (8 header)48 + 864 + 896 + 8
128 + 8192 + 8256 + 8
ms26 2/34050 1/38050 2/3
160213 1/3
ms7.35
10.2913.2419.1225.0036.7648.53
The numbers of packets carried are shown in Fig. 11. Hereit can be seen that the number of packets carried rises asthe packet length decreases. Obviously, more shorterpackets are required to carry the same quantity of speech.The average packet transit times are shown in Fig. 12. Asbefore, the packet transit time is large (over 100 ms) for 25talker pairs. Here, the packet transit time rises for packetlengths greater than 128 + 8 bytes. This is expected, sincelonger packets will obviously take longer to traverse thenetwork. Below 128 + 8 bytes, the packet transit time also
Fig. 1 2 Packet transit time, all talker pairs, varying packet length
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
increases. This rise of transit time in this region is due tothe increased quantity of overhead (8 bytes per packet) ofthe headers of the larger numbers of smaller packets. Thisincreases the trunk utilisation above 90% (as can be seenin Fig. 16) and increases the transit time. Fig. 12 alsoshows a rise in packet transit times below 48 + 8 bytes for20 talker pairs, where the packet transit times is lower thanfor 25 pairs in any case. Referring to Fig. 16 again, onlybelow 48 + 8 bytes does trunk utilisation rise above 80%for 20 talker pairs. It can thus be seen that higher loadingleads to greater packet length below which the headeroverhead leads to sufficiently increased trunk utilisation toincrease the transit time of the shorter packets. The uppertwo curves of Fig. 12 show the same features as Fig. 2 ofReference 10. The minimum transit delay is over 100 ms at128 + 8 bytes (with associated 107 ms packet creationdelay) with a minimum standard deviation of transit timeof over 350 ms (see Fig. 13). This indicates that 25 talkerpairs are more than can be accommodated by the trunkwith reasonable delay (less than 300 ms) with any packetlength. The standard deviation of packet transit times alsorises at short packet length for 25 and 20 talker pairs asthe trunk loading is increased.
The interpacket delays are shown in Fig. 14. These riseapproximately linearly with the speech content of thepackets, and, as before, are dominated by the packet cre-ation times. The results for 15 and 20 talker pairs arealmost indistinguishable. For 25 talker pairs, however, thegraph lies above the other two for all values of packetlength. This suggests that the packets are beginning to takelonger than their predecessors within the same talkspurt.The standard deviation of interpacket delay is shown inFig. 15. As before, this is approximately an order of magni-tude less than the standard deviation of transit time. Theinterpacket delay standard deviation for 20 talker pairs isnot much larger than that for 15 pairs. Both of these are
1.1
I 0.95 0.8I 0.7I 0.6§0.5
</>02
0.10
talker pairs25 ^
2°\32 64 96 128 160 192 224 256d a t a c h a r a c t e r s pe r p a c k e t ( h e a d e r = 8 )
Fig. 13 SD of packet transit time, all talker pairs, varying packet length
Fig. 14 Interpacket delay, all talker pairs, varying packet length
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
much less than that for 25 pairs. The trunk utilisation ofFig. 16 clearly shows the increased quantity of overheaddue to the packet headers as the number of packetsincreases for decreased packet length.
The result of this experiment indicates that the 144kbit/s trunk will not handle 24 conversations of 9600 bit/sbit rate speech with reasonable delays at any packetlength. For 20 talker pairs or less which the trunk can copewith, the packet size should be as short as possible (to giveminimum packet creation delay), consistent with a trunkutilisation which gives acceptable packet transit time (tominimise network delay) and inter-packet delay and itsstandard deviation (to minimise receiver bufferingrequirement). For the network simulated in the experi-ments, this would be between 48 and 128 bytes of speech(plus the packet headers).
Fig. 16 Trunk utilisation, varying packet length, two nodes
3.2.3 Three PSEs, 128 + 8 byte packets, varying talkerload: Fig. 17 shows (for the three networks) the numbersof packets transferred during the experiments. Theexhibited behaviour is seen to be very similar in each case.The packet transit times and their standard deviations areshown in Figs. 18 and 19, respectively. The transit time islow, as before, for less than 22 talker pairs (22 talker pairsoperating between each of the three source-destinationPSE pairs i.e. 66 talker pairs in the whole network). Thetransit time is lowest for all talker loads for the linearnetwork. This is because some of the calls take place overone 288 kbit/s trunk (e.g. between PSEs 1 and 2, orbetween 2 and 3 of Fig. 4c), while a smaller number takeplace over two trunks (between PSEs 1 and 3). Again, withup to 22 talker pairs, the standard deviation of packettransit time is less than 30 ms.
The interpacket delays and their standard deviationsare shown in Figs. 20 and 21, respectively. The inter-packet
ill
delays are very similar up to 25 talker pairs, rise above thistalker load, and are dominated by the 107 ms packetcreation time. The standard deviation of interpacket delaysare again an order of magnitude less than those of thetransit times. The trunk utilisations of Fig. 22 rise linearlywith the number of talker pairs, with little variation
110100
~ 70| 60£50§40J 30H 20a 10
010
Fig. 17nodes
0.9
£0.8
10.7
10 .6o- 0.5
I 0.4oa0.3
10
Fig. 18
2.Or
I 1 5
I 101
"oo
Fig. 19
0.15
- 0.10a,
Ssc
'Z 0.05o
FC-star
Minear
20 30talker pairs (for each of three S-D
Packets transferred, all talker pairs, varying talker load, three
FC
,star
/ /linear
20 30talker pairs (for each of thee S-D )
Packet transit time, all talker pairs, varying talker load
star
/ , linear
10 20 30talker pairs (for each of three S-D pairs)
SD of packet transit time, all talker pairs, varying talker load
FCstarlinear
10 20 30talker pairs (for each of three S-D pairs)
Fig. 20 Interpacket delay, all talker pairs, varying talker load
112
between the three network configurations. The trunk uti-lisations of over 80% for 25 talker pairs and above willobviously lead to the greater delays seen in Figs. 18-21.
The results of this experiment show that networks withcalls operating over more than one trunk (as in the starand linear networks) can, under certain circumstancesprovide reasonable performance (cf. single trunk paths, e.g.the fully connected network) in a packet-switchingenvironment. To do this, and to provide a similar grade ofservice (in delay terms) to the fully connected network,trunks of appropriate capacity were used. If two of thelower speed 144 kbit/s trunks (as used in the fully con-nected network) had been used in a network with two linksper path, the delays would obviously be greater.
^ 0.0150
del
| 0.0100a
| 0.005
0
0
-
/
- ____^£--
FCI
1. stari
J , linear
10 20 30talker pairs (for each of three S-D pairs)
Fig. 21 SD of interpacket delay, all talker pairs, varying talker load
90
80
5s 70
^ 6 0
2 5005 403 30
| 20
~ 100
-
-
-
FC ^ ^ ^
star
^^**^ linear
10 20 30talker pairs (for each of three S-D pairs)
The results of the experiments reported on in this papershow that, under the simulated conditions, there is correla-tion between packet arrivals, and the receiver bufferingrequirement is thus less than that suggested by the spreadof packet transit times. Once the delay for the packets of agiven talkspurt is determined, there is little difference in thedelays experienced by subsequent packets of talkspurt. Thedelays experienced by different talkspurts may, however,differ to a much greater extent. This transit delay coupledwith the packet-creation delay and the receiver bufferingdelay, together with any speech conversion delay, willaffect the total speech delay as experienced by the listener.The timing of the reception of his replies will be affected bythe delay in the reverse direction as well. It would appearthat 1-way and 2-way speech delays are not constant foreach talker, nor for each talkspurt. Clearly, the use of asatellite link (with its 270 ms 1-way delay) in a pathbetween talkers would increase the delay involved in a
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983
packet-switched network even more. For this reason, it isunlikely that packet switching could be used for intercon-tinental voice traffic.
It must be remembered that some simplifications andassumptions were made in the simulation model. Thetalker-activity model was based only on coarse structure,i.e. significant utterances. Fixed-length packets were gener-ated at regular intervals during the talkspurts. This is themain factor which leads to correlation of packet flows andarrivals. If the talker-activity detection had worked on thefine structure, and if packets of different lengths had beengenerated (no filling in of silent periods), it is possible thatmore talkers may have been accommodated in a giventrunk. There is, of course, the problem of the header over-head with these packets. Correlation of packet flows maynot be maintained; thereby requiring longer bufferingtimes. This is particularly important for the proper recon-struction of gaps between syllables, words etc., and thetopic is worthy of further investigation.
Only voice traffic was present in the packet-switchednetworks of the experiments. Data traffic, which is differentin nature, would, of course, be present in an integratedservices network. The data flows are likely to disrupt thecorrelation of speech-packet flow, particularly if the datapackets are (a) much longer, {b) variable in length, or (c)more irregular in entry to the network than the speechpackets. The results of the experiments performed anddescribed in this paper are thus indicative of the best thatcan be achieved.
The effects of both data and voice traffic using the samenetwork should be investigated. Voice and data traffichave different requirements for error control. Speech cantolerate some corruption and still remain intelligible(provided a very low-bit-rate coding scheme is not used)but the speech must not be delayed unduly. Data trafficcan usually tolerate delay, but must not be corrupted. Theerror-control method used in HDLC and X.25 protocolsetc. (which requires retransmission of packets received cor-rectly and discarded after an error, to maintain packetsequence) will obviously delay speech packets. Methods ofdealing with this problem should be investigated. Thesemethods might include the labelling of speech and datapackets and the discarding of any corrupted speechpackets or the discarding of only those whose headers havebeen corrupted (this might require extra error detectionoverhead). Speech and data might be sent, using differentlogical channels (see X.25 etc.) of the same trunk. Eachlogical channel would have the appropriate error and flowcontrol for the type of traffic carried.
Packet switching has been shown to provide acceptablespeech delay performance under some conditions. Furtherwork is necessary to examine the performance of packetswitching in an integrated services network.
Trans., 1964, CS-12, pp. 1-92 ROBERTS, L.G.: 'Computer network developments to achieve
resource sharing'. Proceedings of AFIPS SJCC Conference, May1970, pp. 543-549
3 DA VIES, D.W., and BARBER, D.L.A.: 'Communication networksfor computers' (John Wiley and Sons, London, 1973)
4 KLEINROCK, L.: 'Queueing systems; Vol. II—computer applica-tions (John Wiley and Sons, 1976), pp. 304-314, 422-513
5 MARLOW-MANN, P.D.G.: 'Operational experience and evaluationsof British Post Office experimental packet-switched service', Post Off.Elec. Eng. J., 1979, 72, Apr., pp. 43-49
7 MARTIN, J.: 'System X', Post Off. Elec. Eng. J., 1979, 71, pp. 221-224, (and articles in later issues)
8 BULLINGTON, K.., and FRASER, J.M.:'Engineering aspects ofTASI', Bell Syst. Tech. J., 1959, 38, pp. 353-364
9 KRAUSS, R.M., and BRICKER, P.D.: 'Effects of transmission delayand access delay on efficiency of verbal communications', J. Acoust.Soc. Amer., 1967, 41, pp. 286-292
10 MINOLI, D.: 'Optimal packet length for packet voice communica-tion', IEEE Trans., 1979, COM-27, pp. 607-611
11 MINOLI, D.: 'Issues in packet voice communication', Proc. IEE,1979, 126,(8), pp. 729-740
12 COVIELLO, G.J.: 'Comparative discussion of circuit vs packetswitched voice', IEEE Trans., 1979, COM-27, pp. 1153-1160
13 GRUBER, J.G.: 'Delay related issues in integrated voice and datanetworks', ibid., 1981, COM-29, pp. 786-800
14 GOLD, B.: 'Digital speech networks', Proc. IEEE, 1977, 65, pp. 1636-1658
15 NORWINE, A.C., and MURPHY, O.J.: 'Characteristic time intervalsin telephonic conversations'. Bell Syst. Tech. J., 1938, 17, pp. 281-291
16 BRADY, P.T.: 'A technique for investigating on-off patterns ofspeech', ibid., 1965, 44, pp. 1-22
17 BRADY, P.T.: 'A statistical analysis of on-off patterns in 16 conversa-tions', ibid., 1968, 47, pp. 73-91, Jan 1968
18 BRADY, P.T.: 'A model for generating on-off speech patterns intwo-way conversations', ibid., 1969, 48, pp. 2446-2477
19 AITCHISON, J., and BROWN, J.A.C.: 'The lognormal distribution:With special reference to its uses in economies'. University of Cam-bridge Department of Economics Monographs, Series, No. 5 (CUP,1969)
20 BRADY, P.T.: 'Effects of transmission delay on conversationalbehaviour on echo-free telephone circuits', Bell Svst. Tech. J., 1971,50, pp. 115-134
21 DAVIES, D.W., BARBER, D.L.A., PRICE, W.L., and SOLOMON-IDES, CM.: 'Computer networks and their protocols' (John Wileyand Sons, 1979)
22 CCITT: 'Provisional recommendations, X.3, X.25, X.28 and X.29 onpacket-switched data transmission services' (International Telecom-munications Union, Geneva, 1978)
24 DAHL, O.J., MYRHAUG, B., and NYGAARD, K.: Common baselanguage'. NCC Publ. S-22 (Norwegian Computing Centre, Oslo,1970)
25 GRUBER, J.G.: 'A network interface for frame synchronous speechterminals, and a comparison of measured and calculated speech tem-poral parameters'. 15th HICSS Rec, Honolulu, Hawaii, Jan. 1982, pp.671-680 (see also IEEE Trans., 1982 COM-30, pp. 728-738)
IEE PROCEEDINGS, Vol. 130, Pt. G, No. 4, AUGUST 1983 113
