Analysis of the Delay and Jitter of Voice Traffic Over the Internet

Transcription

1 Analyss of the Delay and Jtter of Voce Traffc Over the Internet Mansour J. Karam, Fouad A. Tobag Abstract In the future, voce communcaton s expected to mgrate from the Publc Swtched Telephone Network (PSTN) to the Internet. Because of the partcular characterstcs (low volume and burstness) and strngent delay and loss requrements of voce traffc, t s mportant to separate voce traffc from other traffc n the network by provdng t wth a separate queue. In ths study, we conduct a thorough assessment of voce delay n ths context. We conclude that Prorty Queung s the most approprate schedulng scheme for the handlng of voce traffc, whle preempton of non-voce packets s strongly recommended for sub-10 Mbt/s lnks. We also fnd that per-connecton custom packetzaton s n most cases futle,.e. one packet sze allows a good compromse between an adequate end-to-end delay and an effcent bandwdth utlzaton for voce traffc. I. INTRODUCTION The Internet wll become a ubqutous nfrastructure, used by numerous applcatons havng varous requrements and that generate traffc that has dfferent characterstcs: n partcular, web-based data applcatons, vdeo applcatons and voce applcatons. Voce applcatons are expected to mgrate from the Publc Swtch Telephone Network (PSTN) that servces them today to the Internet. Owng to constant mprovements over the years, tradtonal voce communcaton over the PSTN s today characterzed by what s referred to as toll qualty, that s low delay, hgh avalablty and adequate voce qualty. For the Internet to compete wth the PSTN, t should provde the same level of qualty, whch mples strngent delay, loss and relablty requrements on voce communcaton. In terms of traffc characterstcs, voce streams have low data rates (n the order of tens of Kbt/s) and exhbt low burstness. Because of these strngent requrements and partcular characterstcs, voce traffc should be treated dfferently than other traffc n the network. In fact, measurements on the Internet [28] as well as smulaton studes [22] have shown that mxng voce traffc wth both tradtonal TCP data traffc and UDP VBR vdeo traffc can lead to ether low average lnk utlzaton f delay requrements are met, or larger than desred delay for voce. Accordngly, allowng a mxture of voce wth other traffc can lead to the need of complex admsson control polces f the end-to-end delay requrements for voce were to be satsfed [22]. Hence, we assume n ths study that voce traffc s separated from other traffc n the network by provdng t wth ts own queue. That s, voce could be provded wth a separate lnk, or a separate crcut usng, for example Mult-protocol Lambda Swtchng [1]. If other traffc s flowng on the lnk, then voce traffc could be servced usng Prorty Queung (PQ), and gven the hghest prorty over all other traffc. In LANs, such a treatment s suggested n the context of the IEEE 802.1p extenson to the IEEE 802.1D standard [18]; at the Internet scale, hgh prorty can be provded to voce Mansour Karam s wth Speedtrak Communcatons, San Mateo, CA Emal: mans@speedtrak.com. Fouad Tobag s wth the Computer Systems Laboratory, Department of Electrcal Engneerng, Stanford Unversty, Stanford, CA 94305, USA. Emal: tobag@stanford.edu traffc n the Dfferentated Servces framework [2], by means of mappng voce traffc to the Expedted Forwardng Per-hop Behavor (EF PHB) [19], and gvng hgh prorty to EF traffc relatve to other traffc n the network. Fnally, voce traffc could be gven ts far share of the lnk usng Weghted Round Robn (WRR) [24] or comparable schemes ([14], [30], [32]). In ths paper, we am to demonstrate that n ths context, the requrements of voce traffc can be attaned usng smple mechansms, both n terms of schedulng and packetzaton. In ths respect, we start wth descrbng the partcularty of voce traffc, n terms of characterstcs, requrements and delay components. Snglng out queung delay as the only source of jtter, we present the methodology used to quantfy t. We come up wth approprate models for voce delay, justfy the ntuton behnd ther choce and show ther accuracy by comparng the resultng delay dstrbutons to those obtaned through network smulaton. Usng these models, we show that schedulng voce traffc usng PQ and packetzng voce streams usng a fxed sze packetzaton scheme lead to an adequate handlng for voce traffc n the Internet. The paper s organzed as follows: we start wth a revew of pror work (Secton II). In Secton III, we descrbe the partcularty of voce traffc n terms of characterstcs, requrements and delay components. In Secton IV, we descrbe the models used for quantfyng the network delay and jtter ncurred by voce traffc n the Internet. We thereafter present the results of the analyss. In Secton V, we compare the delay obtaned for voce traffc usng dfferent schedulng schemes. By the same token, we also nvestgate the effect of varous parameters on network delay and jtter. In Secton VI, we comment on the choce of packet sze for an adequate handlng of voce traffc. Fnally, we end n Secton VII wth a summary of the results and some concludng remarks. II. PRIOR WORK Snce the concept of a packet network that supports both voce and data traffc emerged n the early eghtes, the work on packetzed voce has been extensve. Queung models were developed, used and compared to understand the queung behavor of voce traffc n a packet network. In partcular, the superposton of multple perodc streams, whch consttutes the most ntutve model for packetzed voce was presented and analyzed analytcally n [9]. The queung model obtaned s denoted. 1 [8] and [21] used to derve the ds- In, the nput process conssts of the superposton of a fxed number of perodc streams, and each stream s characterzed by a determnstc servce tme. The dstrbuton of the queung delay ncurred by a packet n a stream s obtaned, assumng dfferent nstances of the same experment, where for each nstance the phase of the varous stream wth respect to each other s chosen at random.

2 trbuton of the queung delay that s ncurred by voce traffc, whle [23] also derved the correspondng buffer dstrbuton. [9], [21], and [29] compared the delay results obtaned usng to those pertanng to the smpler queung model, whch was commonly used n practce to estmate the szng of real systems. These studes found that sgnfcantly over-estmates the delays ncurred on packetzed voce when the utlzaton on the lnk s hgh and the number of streams multplexed on the lnk low. On the other hand, [9], [21], and [29] found that n the case of lghtly utlzed hgh speed lnks, where the utlzaton s low and the number of streams multplexed exceeds 100, and yeld smlar results. [31] studed the effect of Speech Actvty Detecton (SAD) (that s, slence suppresson) on voce delay: t was found that the ncrease n traffc varablty that results from the ncluson of SAD n the encodng process hnders the advantage that s obtaned from the reducton n average utlzaton 2. Fnally, [26] used the and models to derve onehop queung delay, and convoluton over many hops to derve mult-hop queung delay for voce traffc. Today, as the Dfferentated Servces archtecture [2] has ganed n popularty, new studes have emerged, amng at specfyng a servce based on the EF PHB [19] that would be approprate for the handlng of voce traffc. PQ has been proposed by some as a smple and adequate schedulng scheme n ths context [20], but refuted by others [7] because of the sgnfcant queung delay varatons (.e., jtter) that could be ncurred as a result of the resdual transmsson tme of lower prorty packets. In fact, [7] shows that the worst-case queung delay ncurred by voce traffc ncreases exponentally wth the number of hops traveled, as a result of the ncrease n traffc burstness wth the number of hops. Conversely, [16] and [27] show that settlng for a statstcal guarantee (.e., acceptng a small porton of lost traffc) allows a sgnfcant reducton n delay, and suggest that PQ s ndeed adequate for the handlng of voce traffc. However, the results obtaned are nconclusve, as the mult-hop scenaro used n both studes underestmates the ncrease n traffc burstness as the number of hops n the path of a voce stream ncreases. III. PARTICULARITY OF VOICE TRAFFIC A. Voce Traffc Characterstcs and Requrements Voce Characterstcs. The tradtonal voce encoder s G.711 (and ts varants [10], [11]), whch uses Pulse Code Modulaton (PCM) to generate 8 bts samples per 125 mcroseconds, leadng to a rate of 64 Kbt/s. In the last decade, new voce encodng schemes have been developed, whch use Code Book Excted Lnear Predcton (CELP) technques, leadng to drastc rate reductons at the expense of addtonal encodng delay: 8 Kbt/s for G.729A [13], 5.33 Kbt/s for G [12]. Takng nto account the headers that correspond to each of the protocol layers, the rate of the packetzed voce stream remans n the order of tens of Klobts per second, whch s much lower than the data rates that correspond to typcal vdeo and data traffc. In addton, speech conssts of an alternaton of talk-spurts and slence More specfcally, [31] found that the queung delay resultng from an model n whch the rate of the ncomng (Posson) process s set equal to the average ncomng rate of voce traffc heavly underestmates the delay ncurred by voce traffc n the network. Fg. 1. End-to-end delay components for voce traffc. perods. Slence suppresson at the source takes advantage of ths fact, leadng to a substantal reducton of the average rate at the expense of ncreased varablty. Voce Requrements. Voce requrements are strngent: tollqualty real-tme communcaton s needed, whch lmts the maxmal tolerable round-trp delay to ms; that s, oneway delay must be n the range of ms for adequate performance. In addton, jtter should be small enough (.e. 50 ms at most) so that playback at the recever remans smooth. On the other hand, the tolerable packet loss s small. In fact, snce packet loss n the Internet s correlated [3], f packet loss were to occur, the number of contguous packets whch are lost s usually larger than one. Hence, the duraton of the correspondng porton of voce bt-stream that s lost (whch we refer to as a clp ) can easly exceed 60 ms even when a smaller packet formaton tme s used. [17] shows through subjectve testng that clps exceedng 60 ms affect the ntellgblty of the receved speech. For ths reason, for toll-qualty communcaton, must be set to a relatvely low value (.e., ) to nsure that such clps occur nfrequently. Takng advantage of as opposed to conductng a worst case analyss s necessary, snce the later would lead to delay results whch are extremely pessmstc (e.g., [7]), and that do not apply to realstc stuatons. Consequently, n the remander of ths study, we measure delay by the percentle, where. B. End-to-end Delay Components of Voce Packets End-to-end delay conssts of the delay ncurred by the voce sgnal from the nstant t s produced by the speaker untl t s heard by the lstener at the destnaton (see Fgure 1): the analog sgnal s frst encoded, ncurrng an encodng delay!, whch n turn conssts of the sum of the frame sze #, the look-ahead delay$, and the processng delay% (see Table I). In general, the lower the rate of the encoded bt-stream&, the larger the frame sze, look-ahead delay and processng delay, and consequently!. The encoded bt-stream thus generated s then packetzed, ncurrng a packetzaton delay ('! ), functon of the number of frames* ncluded n one packet,.e. '! ) * #. Voce packets are then transmtted on the network, ncurrng transmsson delay +, queueng delay, and propagaton delay- at

3 TABLE I FRAME SIZE, LOOK-AHEAD DELAY AND PROCESSING DELAY FOR G.711, G.729A AND G Encodng G.711 G.729A G scheme (Nomnal bt (64 Kbt/s) (8 Kbt/s) (5.3/6.4 Kbt/s) rate ) Frame sze 125 s 10ms 30ms Look-ahead 0 5ms 7.5ms Processng delay Neglgble Less than 10ms Less than 30ms Fg. 2. General network Topology. each hop n the path from the source to the destnaton. Propagaton depends on the dstance between the source and the recever 3. At the recever, packets are delayed n a playback buffer ncurrng a playback delay '. De-packetzaton s then performed, and the reconstructed encoded bt-stream! s decoded at the destnaton ncurrng a decodng delay. We consder processng to be fast enough so that processng delay at both the source and the recever (%! and ) are gnored. Therefore, denotng the formaton tme + * #, we get + $ +, - ' (1) From Equaton (1), t s clear that for a gven voce connecton, the only random component of voce delay (that s, the only source of jtter) conssts of queung delay n the network,,,. The playback delay ' nsures that most of the packets transmtted are avalable the nstant they have to be handed to the decoder. Assumng that, represents the maxmum queung delay percentle ncurred n the network, the recever must delay the frst packet of a voce stream by a full,, 4.e. ', ; f, n addton, that packet has already ncurred n the network a queung delay equal to,, then the end-to-end delay budget equaton becomes + $ + -, (2) Equaton (2) clearly shows the mportance of queung delay, formaton tme and propagaton delay. In partcular, t can be used to estmate the maxmum jtter, that can be tolerated for a gven connecton; for example, f G.729A s used to encode the voce stream, a formaton tme+ ms s used to packetze the encoded bt stream, the total transmsson and propagaton delay on the path are equal to 5 and 40 ms, respectvely, then for ms,, ms. Transmsson tme beng neglgble most of the tme (except on very For calls wthn a gven local area, propagaton delay s neglgble. For ntracontnental calls wthn the Unted States (e.g., San Francsco to Boston), the propagaton delay s n the order of 30 ms whereas nter-contnental calls result n propagaton delays rangng from 50 ms (e.g., San Francsco to Pars) to 100 ms! (e.g., San Francsco to Hong-Kong). Even though RTP provdes a sequence number and tme-stamp for each packet, source-recever synchronzaton s not supported. Hence, the source and recever are typcally non-synchronzed, so the recever can t determne the amount of jtter already ncurred by the frst packet receved. slow lnks), Equaton (2) clearly shows the mportance of queung delay, formaton tme and propagaton delay, whch are the three components we focus on n ths paper. IV. MODEL FOR QUEUING DELAY IN THE NETWORK In ths secton, we am to dentfy the sources of queung delay ncurred by voce traffc and derve models that captures the effects of nterest. We valdate our models usng network smulaton. The general network topology consdered s shown n Fgure 2; t conssts of a successon of hops n tandem, whch represents the path followed by a voce stream from the source to the recever. To be conservatve, we consder the pessmstc scenaro n whch the traffc that s nterferng wth the tagged stream s njected at each hop, ndependently of the other hops n the path. Also, we assume the general case n whch the nterferng traffc can be generated by a source many hops away from the swtch at whch the nterference wth the target stream takes place (as opposed to the source of the nterferng traffc beng drectly attached to the swtch n queston). Dong so, the varablty of the nterferng voce traffc gets larger as t travels through multple hops before t s made to nterfere wth the target stream, leadng, n turn to a larger queung delay. Ths scenaro s more reflectve of realstc traffc condtons. We consder a number of lnk speeds, rangng from 384 Kbt/s to 45 Mbt/s, ncludng T1, 10Base-T and T3 lnks. As descrbed n Secton I, we sheld voce traffc from other traffc n the network by provdng t wth a separate queue. We consder that Slence Suppresson s mplemented by the voce encoder, yeldng a decrease n the total voce load. Accordngly, the model used for voce traffc conssts of an ON/OFF pattern, modelng the successon of talkspurts and slence perods generated by the voce encoder. As descrbed n [8], talkspurts and slence perods alternate accordng to a two-state Markov chan. In our experments, we consder average perods of talk-spurt and slence equal to 1.65 and 1.35 seconds, respectvely ([8]). We dentfy two sources of voce delay jtter (see Fgure 3): 1. Queueng delay behnd voce packets n the same queue: ths component depends on the varablty of the voce traffc pattern whch, at the source orgnates from slence suppresson. In case the lnk (or crcut) on whch voce travels s not shared wth other traffc, then a gven voce stream s not affected as t travels through consecutve hops from the source to the recever, that s ts varablty remans the same at each hop n the path. In

4 Fg. 3. The two source of jtter for voce traffc when provded wth a separate queue. ths context, the nterferng traffc at a gven hop can be consdered to be generated by sources drectly attached to the swtch n queston. 2. In case voce traffc shares the lnk wth other traffc, t then ncurs the resdual transmsson tme of non-voce packets. For example, n case non-preemptve PQ s used, then voce packets that arrve to the queue durng the transmsson of a lower prorty packet wll be delayed untl the end of the ongong transmsson of that packet. Also, as voce packets correspondng to a gven stream ncur such a delay, ther nter-arrval tme s modfed: as a result, voce traffc varablty ncreases wth the number of hops, n turn leadng to an ncrease n the queung delay ncurred n the voce queue. Accordngly, the jtter experenced clearly depends on whether voce traffc shares the lnk (or crcut) wth other traffc. In the followng secton, we study each case ndependently. A. Voce Alone on the Lnk or Crcut Wthout slence suppresson, the encoder produces frames at regular ntervals #, whle packets are generated at regular ntervals +. That s, each voce stream s perodc wth perod +. Also, snce packets are of equal sze, then the servce tme, where denotes the header sze n bytes. The queung model that s characterzed by an nput process that conssts of a superposton of ndependent perodc sources, and by a determnstc servce tme s, and has receved extensve attenton n the past 5. (See Secton II.) In our case, slence suppresson s mplemented, so the voce traffc load s reduced and ts varablty ncreased. We have smulated a number of dfferent scenaros usng the topology shown n Fgure 2. In Fgure 4, we show the complementary queung delay dstrbuton and compare t to the complementary queung delay dstrbuton that results from a model. The plot shows that the delay dstrbuton obtaned usng the model has a fast decayng tal, of packets s determnstc, equal to & + whch s expected gven the low level of burstness n the traffc. Clearly, when slence suppresson s mplemented, then the resultng complementary delay dstrbuton s bounded by that obtaned wth for low delay values. However, the tal of the dstrbuton s manly affected by the perods n whch, the only source of randomness orgnates from the Note that n random phase of each of the perodc streams wthn a perod. Fg. 4. Model versus smulaton results. all voce streams are n ther talk-spurt. Thus, the tal of both dstrbutons match very closely, and both models lead to nterchangeable delay results for the tolerable loss rate consdered n ths study. Ths phenomenon agrees wth prevous fndngs showng that the varablty that entals from slence suppresson ncreases the ncurred delay percentles, thus reducng the advantage obtaned through an ncrease n the multplexng gan [29] [31]. (See Secton II.) Havng observed ths result for a large range of scenaros, we conclude that the model approxmates well the delay results obtaned n case voce s transmtted alone on the lnk. Snce hops are ndependent, then the total delay ncurred by voce packets travelng through a gven number of hops s dstrbuted accordng to the -fold convoluton of the one hop delay dstrbuton. B. Voce Sharng the Lnk wth Other Traffc As descrbed above, n case voce shares the lnk wth other traffc, t ncurs the resdual transmsson tme of packets transmtted over lower prorty queues. We treat the case of PQ frst; we then extend the results to WRR. We frst characterze the added varablty that results from the perturbatons on the hgh prorty queue from packets belongng to the lower prorty queues. We show that the resultng varablty s bounded by that nherent to a Posson process. We then consder the voce nput process to be Posson, and fnd the resultng dstrbuton of queung delay ncurred by voce traffc. For a gven hop, we have ) ) ) ) ) ) where ) ), ) ) represent the tme nterval separatng two consecutve packets * and * belongng to the same stream before ther admttance to the ) queue, and after ther transmsson on the lnk, respectvely; represents the perturbaton ntroduced by lower prorty packets over one hop, and s essentally equal to ther resdual transmsson tme. Assumng ) that lower prorty packets are all full sze MTU packets, s unformly dstrbuted, rangng from 0 to +, where + denotes the transmsson ) ) tme of a full sze MTU packet. Thus, the perturbaton ) s dstrbuted accordng to a symmetrc, trangular dstrbuton rangng from +

5 [ X F F Fg. 5. for! $# and % (exponental, wth mean ) for varous number of hops and lnk speeds. to +, wth varance &$' )( /. Furthermore, the perturbaton ncurred by the voce packets belongng to a stream travelng a gven number of hops s the sum of the perturbatons ncurred at each hop,. Snce hops are ndependent, then the dstrbuton correspondng to the total perturbaton s dstrbuted accordng to the -fold 3 convoluton of the dstrbuton of. From the central lmt theorem, we know that as ncreases, where ;: =<?> + 7@ 1/A represents a normal dstrbuton wth mean 0 and standard devaton &)1 + 7@ 1/. The convergence of 0!1 to B8;: <?> +,@ 1/A can be shown to be fast (wthn 10 hops), whereas the tal of the complementary dstrbuton functon of the lmtng normal dstrbuton always bounds that of 0 1. Therefore, B8;: < C> +,@ 1/A consttutes a close approxmaton to 0!1. Usng 021 to compute 1) the nter-arrval dstrbuton of packet szes D 1 1) IH for a voceih stream travelng through hops, we get#.e + GF,.e. D #KJML J ( 1 s normal wth mean + and standard devaton+ N@ 1O. 6 *,+.- Note that the standard devaton of the nter-arrval process grows wth the square root of the number of hops n the path of the voce stream; that s, ncreasng the number of hops only results n a contaned ncrease n varablty. Also, the tal of the MQ dstrbuton of D 1 decays wth P ( ; ntutvely, ths shows that the resultng burstness n the nput process s lower than the burstness of MQ a Posson arrval process, whch dstrbuton tal decays wth P. To quantfy ths observaton, we compare the dstrbuton of D 1 to that of an exponental dstrbuton R P UT L ). In Fgure 5, we plot the probablty that the dfference between the nter-arrval perod and + exceeds a gven perturbaton W, a good measure of wth mean+, (.e#s IH + the stream varablty. As can be seen from the fgure, the var- Snce Y and UY are both normal dstrbutons wth mean 0 and varance Z, then Y Y s a normal dstrbuton wth mean 0 and varance Z ]\ _^a` Z. V Q ablty of D 1 only exceeds that of R when lnks are extremely slow (384Kbt/s). For T1 lnks, the varablty of D 1 s the lowest for as much as a 19 hops path (that s, for all practcal cases). Hence, f PQ s used, wth voce traffc gven the hghest prorty over other traffc n the lnk, we can use a Posson nterarrval process of mean + to bound the varablty that results from the effect of lower prorty non-voce packets. One sde beneft n so dong s to render the dstrbuton of nter-arrval tme ndependent of the hop number n the path. We now fnd the dstrbuton of the queung delay ncurred by a voce stream as t traverses a gven hop; we use n that end the treatment of prorty functons found n [25]: frst, we derve the Laplace transform for the watng tme for voce packet n a two prorty system where the hgh (voce) and the low (non-voce) queues are both fed by a Posson nput process, and where the servce tme for packets n both queues s determnstc, equal to and, respectvely. We get ' bdc fe hg j me mnpo < ljc ljc fe '_k ' k A (3) 3 q where, represent the average utlzaton of the hgh prorty ' voce queue and lower prorty non-voce queues, respectvely, represent the total average utlzaton of the lnk, ' l c and fe k v T.r sut > w x> t s the Laplace transform of the resdual lfe of the servce tme for packets n each queue. To be conservatve, we assume that the traffc n the lower prorty queue uses up all of the bandwdth that s unused by voce traffc; that s, we set to 1; thus, Equaton (3) becomes jï < bdc fe zy o ljc Ie k A ljc k ' 3 q that s, ~} ]}, where k ' ~} ]} represents the watng tme obtaned from an queung system, whereas represents the resdual transmsson tme of packets n the k ' lower prorty queue. Thus the delay ncurred by voce packets s smply the sum of an derved watng tme and the resdual transmsson tme of lower prorty packets. Also, snce the model s ndependent of the hop number, and snce hops are ndependent, then the dstrbuton of the queung delay ncurred by a voce stream that travels through hops s smply the -fold convoluton of the queueng delay ncurred by a voce stream over one ndvdual hop. In order to verfy the proposed model, we have smulated a number of scenaros that are based on the mult-hop topology descrbed above. In partcular, we consder a herarchcal topology n whch the nterferng traffc at each of the hops vsted by the target voce stream s assumed to have traveled the same number of hops as the target voce stream. In Fgure 4, we plot for one such scenaro the complementary dstrbuton of the queung delay ncurred by voce packets belongng to the target stream as obtaned from smulaton, and compare t to that obtaned by the model presented above. Consstent wth our assumpton, we consder n the smulatons that the lower prorty packets are all MTU-szed, and that the lower-prorty queue never emptes. As can be seen from the plots, the dstrbuton obtaned from the model closely bounds the dstrbuton.{ fe

6 obtaned from the smulatons; ths fact has been observed for a number of dfferent scenaros based on the general topology. When WRR s used nstead of PQ, two addtonal effects come nto b play: on one hand, gven that the voce traffc s gven a share of the total bandwdth, then the actual bandwdth that! s avalable for voce traffc, b could be as low as (that s, n case the remanng bandwdth s fully utlzed by other traffc). In addton,! because schedulng s done at the packet granularty, b vares around, whch ncreases traffc varablty and consequently delay percentles. On the other hand, a gven voce packet could ncur the transmsson tme of more than one non-voce packet, dependng on the number of queues that contend for the lnk. We ntend to obtan delay results that correspond to a lower bound on the maxmum delay ncurred by voce traffc when servced by a WRR scheduler. Hence, we assume b (1) that the avalable bandwdth for voce traffc s equal to at all tmes, and (2) that voce packets ncur a resdual transmsson tme correspondng to one non-voce packet. ^ ms, 50% utlzaton) Fg. 6. CCDF of Queung Delay (T1, G.729A, for 1, 2 or 5 hops. Two confguratons are compared: ether voce s transmtted alone on the lnk, or gven hgh prorty over other traffc flowng on the lnk. V. CHOICE OF SCHEDULING SCHEME Usng the models developed n Secton IV for voce delay, we now propose to compare dfferent schedulng schemes that can be used to servce the voce queue. We frst nvestgate n Secton V-A n detal the effect of the resdual transmsson tme of non-voce packets on voce queueng delay. By the same token, we consder the effect of varous network parameters (lnk utlzaton, number of hops and lnk bandwdth) on voce delay. We then compare n Secton V-B PQ to competng schemes, and establsh ts adequacy n support of voce traffc. A. Resdual Transmsson Tme of Non-Voce packets In case voce traffc s provded wth a separate lnk (or crcut), t s completely shelded from other traffc on the network. If voce traffc shares the lnk wth other traffc nstead, wth PQ used to schedule packets, then as found n Secton IV, the ncrease n queung delay caused by the resdual transmsson tme of non-voce packets s two-fold: on one hand, the queung delay ~} ]} s sgnfcantly larger than t } }. Except for lghtly utlzed hgh speed lnks (see Secton II), leads to sgnfcantly larger delay values as compared to. 7 On the other hand, the resdual transmsson tme of lower prorty packets s typcally much larger than ~} ]} ; n fact, consderng a typcal MTU of 1500 bytes, the transmsson tme of an MTU szed packet s around 16 tmes larger than that of a typcal voce packet 8. Number of hops and lnk utlzaton. In Fgure 6, we plot the complementary dstrbuton of queung for both confguratons n the case of G.729A, for one, 2 and 5 hops, over T1 lnks. The frst observaton s that n all cases, the tal of the delay dstrbuton s stll fast-decayng, whch confrms the fact that the delay percentle s ndeed much lower than the maxmum delay that could be acheved n the network. However, the tal of the complementary delay dstrbuton wdens sgnfcantly when voce shares the lnk wth other traffc. For In partcular, as the average aggregate utlzaton approaches 1, t never exceeds, whereas ncreases wthout bound. That s, a packet generated by G.729A, packetzed usng = 30 ms and to whch a 46 bytes header s appended, for a total of 76 bytes. ^ Fg. 7. CCDF of Queung Delay (T1, G.729A, ms, 5 hops) wth voce traffc totalng a maxmum utlzaton of voce 10, 50 and 75%. Two confguratons are compared: ether voce s transmtted alone on the lnk, or gven hgh prorty over other traffc flowng on the lnk. a tolerable loss rate set to, the queung delay percentle ncurred by voce flowng alone on the lnk does not exceed 6 ms even f the path conssts of fve T1 lnks; conversely, when voce traffc s gven prorty over other traffc on the lnk, then voce delay s close to 20 ms as soon as two T1 lnks are traversed. In fact, queung delay becomes very senstve to the number of hops (e.g. more than 40 ms ncurred for 5 T1 lnks), whch s expected, snce for, the percentle of the resdual transmsson tme over hops s close to tmes +. Clearly, n ths case, the contrbuton of the resdual transmsson tme of lower prorty packets to queung delay s large. Smlar observatons apply to other lnk speeds (10 Mbt/s, 45 Mbt/s, etc.). In Fgure 7 we plot the complementary dstrbuton functon of queung delay for the same two confguratons n the case of G.729A for 5 T1 hops, and for voce traffc usng up to 10%, 50% and 75% of the lnk bandwdth 9. The fgure reveal that when voce traffc s gven prorty over When measurng lnk utlzaton, we assume a gnore the effect of slence suppresson. model, and hence

7 Fg. 8. CCDF of queung delay (G.729A, 5 and 10 hops, and for T1, 10Base-T and T3 lnks. ^ ms, 50% utlzaton) for 1, other traffc on the lnk, the ncrease n queung delay that results from an ncrease n voce utlzaton from 10 to 75% becomes less sgnfcant, as compared to the total queung delay ncurred. That s, when voce shares the lnk wth other traffc, the characterstcs of voce traffc become less mportant, whle the resdual transmsson tme of non-voce traffc becomes the determnng factor. Avalable Bandwdth. In general, the effect of bandwdth on queung delay s well known n queung theory, as average queung delay for a gven utlzaton s shown to be nversely proportonal to the avalable bandwdth for most queung systems of nterest. Smlarly, delay percentles (as consdered n ths study) decrease wth an ncrease n the avalable bandwdth 10. As queung delay decreases wth the avalable bandwdth, other delay components do not vary: n partcular, propagaton delay, formaton tme and look-ahead delay are ncurred by all voce packets n the stream; therefore, when bandwdth s large enough, worryng about queung delay becomes futle. In Fgure 8, we plot the complementary delay dstrbuton of voce traffc for dfferent values for the number of hops n the path and the avalable bandwdth. One can see from the graph that the queung delay percentle ncurred on voce packets gong through 10 T3 hops remans lower than 3 ms whle the delay ncurred over one and 5 T1 hops already exceeds 10 and 40 ms, respectvely, whch s on the order of both the formaton tme and the propagaton delay, and consttutes a large proporton of the tolerable end-to-end delay ms. B. PQ Versus Other Schedulng Schemes The Secton V-A above, we showed that PQ leads to adequate delays when the avalable bandwdth s large. In ths secton, we look more closely at the approprate choce of schedulng scheme, and compare n that respect PQ, WRR, and the provson of a separate crcut for voce traffc. We plot n Fgure 9 the complementary delay dstrbuton for voce traffc when For, delay percentles are nversely proportonal to the avalable bandwdth for a gven utlzaton (that s, for a proportonally larger number of streams supported). For, queung delay percentle also decreases when the avalable bandwdth s ncreased, but the two are not exactly nversely proportonal. Fg. 9. Complementary delay dstrbuton for voce traffc served usng ether PQ, WRR or when provded wth a separate crcut. (In the case of WRR, the results show a lower bound on the complementary dstrbuton of voce delay for two shares allocated for voce traffc, 1.5 and 10 Mbt/s respectvely.) servced usng preemptve and non-preemptve PQ, and lower bounds of the complementary dstrbuton b of voce traffc when servced usng WRR for two values of b (1.5 Mbt/s and 10 Mbt/s, respectvely). Note that when s set to a value that s close to the maxmum voce load (1.5Mbt/s versus 1.1Mbt/s), the delay ncurred b by voce traffc exceeds 10 ms for 5 T3 hops; however, as s ncreased to 10 Mbt/s, then the queung delay ncurred by voce traffc decreases to less then 2 ms, whch s very close to the value obtaned wth non-preemptve PQ. b In fact, as long as the rato of the maxmum voce load to s kept low (below 20%), the delay ncurred wth WRR converges quckly to that b obtaned wth non-preemptve PQ. Also, note that even though s much larger than the maxmum voce load, resources are not wasted snce resdual bandwdth that s unused by voce traffc can be shared by traffc n the other queues. In addton, WRR has the well-known advantage of makng sure that no traffc from a gven queue starves traffc from other queues. However, snce voce traffc wll reman a small porton of Internet traffc (10% of the total Internet traffc expected wthn 5 years), we suspect that, n most stuatons, no specal precauton wll be needed to prevent voce traffc from starvng other traffc. Also, note that WRR can lead to delays that are much larger than those shown n ths fgure: n partcular, when multple queues share the lnk, then voce packet could ncur the delay pertanng to more than one MTU-szed packets. For ths reason, WRR s not deal for the handlng of voce traffc. We now compare non-preemptve PQ to the provson of a crcut for voce traffc. As shown n Fgure 9, the penalty that results from havng voce traffc ncur the resdual transmsson b tme of other traffc s such that carvng out a crcut of Mbt/s for voce traffc on a 45 Mbt/s lnk (thus sheldng t totally from other traffc on the lnk) leads to lower delays than gvng prorty to voce traffc over the entre 45 Mbt/s bandwdth. However, contrarly b to WRR, the dfference between the actual voce load and s wasted n ths case. That s, compared to PQ, lower delays can be obtaned by ether wastng resources (whch s undesrable), or by preemptng the transmsson of non-voce packets. To acheve a delay for voce traffc that s lower than that offered by a non-preemptve PQ

8 ^ scheduler, yet avodng the waste n bandwdth that results from provdng voce wth a separate crcut, a PQ scheduler can be made to preempt the transmsson of lower prorty packets, as has been proposed for low speed lnks 11 [4]. In ths case, the delays obtaned become as low as f voce traffc were flowng alone on the lnk. As shown n Fgure 9, the use of preempton reduces the delay ncurred by voce traffc across 5 hops over T1 lnks to less than 0.2 ms. In general, the results show that wth preemptve PQ, the queung delay of voce traffc flowng across 5 hops over T1 lnks does not exceed 11 ms for a lnk utlzaton as hgh as 75%. Therefore preemptve PQ s ndeed the most approprate schedulng scheme for handlng voce traffc over sub-10 Mbt/s lnks, whle non-preemptve PQ s adequate (that s, leads to a queung delay that s lower than 10 ms) when the lnk bandwdth exceeds 10 Mbt/s. 12 Fg. 10. versus formaton tme for G.729A and G.711 VI. CHOICE OF PACKETIZATION SCHEME In ths secton, we look nto the effect of packet sze on voce delay and bandwdth utlzaton. We show that n most cases, choosng a packet sze dynamcally on a per-connecton bass only provdes a modest beneft as compared to usng a fxed packet sze for all connectons. Headers correspondng to the varous layers of the protocol stack are appended to voce packets before they can be transmtted on the network 13. Denotng & the rate of the encoded bt stream, and : the rate of the packetzed bt stream, we have : & V. In Fgure 10, we plot the rato as a functon of + for both G.729A and G.711 over pont-to-pont lnks. Clearly, the lower the rate of the encoded stream, the larger the overhead, and hence the larger the rato. When + 10 ms, : & for G.729A; ncreasng the formaton tme to 30 ms already decreases ths rato to, whle a formaton tme as hgh as 100 ms s needed to decrease the overhead to 50%. Therefore, there s an ncentve to use the largest formaton tme possble gven the maxmum tolerable delay, the propagaton delay - - and the queung delay,. In ths secton, we ntent to nvestgate whether dynamc packetzaton (usng the largest formaton tme as descrbed above) s benefcal. We show that n most cases, choosng a packet dynamcally on a per-connecton bass only provdes a modest beneft as compared to usng a fxed packet sze for all connectons. We start wth the end-to-end delay budget equaton derved n Secton III-B (Equaton 2). Ignorng the transmsson tme In ths case, when a voce packet reaches the head of the hgh prorty queue whle a lower prorty packet s beng transmtted, then the transmsson of the data packet s nterrupted to allow the transmsson of the voce packet; only when the hgh-prorty voce queue s empty agan does the transmsson of the data packet resume. Ths result contradcts the fndngs of [7], n whch PQ was shown to lead to extremely large delay values n the network. The reason for the dscrepancy les n the choce of the measure used n [7], the worst-case delay. Ths once agan confrms the mportance of takng advantage of the tolerable packet loss n the Internet. A 12 byte RTP header, an 8 byte UDP header, a 20 byte IP header, and ether a 6 byte data-lnk (e.g., PPP or HDLC) or a 29 byte MAC header dependng on whether the voce packets are crossng a pont-to-pont ^ lnk or an Ethernet LAN, respectvely, for a total header sze or 69 bytes, respectvely. Fg. 11. Maxmum number of streams versus on T1 and T3 lnks, for a 5 hops path and a tolerable end-to-end delay ms. The propagaton delay s consdered neglgble for all voce streams sharng the lnk. (whch, as stated n Secton III-B s, n general neglgble compared to other the other delay components), we get, for a gven stream, +, - $ (4) In case dfferent encodng and packetzaton schemes are used for each voce stream beng served by a network node,+ and$ can dffer across them. For smplcty, we consder that all voce streams use the same encoder. We frst consder G.729A (that s, $ ms), and then comment on both G.711 and G Also, n a typcal network node servng voce traffc, dfferent streams are characterzed by travel paths havng dfferent hop counts and dfferent propagaton delays. Before nvestgatng such a scenaro, we consder the smple scenaro n whch all streams on a lnk travel the same number of hops and have a zero propagaton delay. In other words, gven a maxmum tolerable end-to-end delay, the end-to-end delay requrement for voce becomes +, $ For ths scenaro, we plot n Fgure 11 the maxmum number

9 TABLE II DISTRIBUTION OF PROPAGATION DELAY FOR VARIOUS GEOGRAPHIC LOCATIONS. Propagaton (ms) Delay Dstrbuton for local areas Dstrbuton for wde areas Dstrbuton for transoceanc lnks of streams as a functon of + on T1 and T3 lnks for a maxmum tolerable end-to-end delay ms. The number of streams ncreases at frst wth +, as a result of the reducton n bandwdth requrement per stream; however, as the formaton tme approaches $, the total utlzaton on the lnk must decrease to keep queueng delay low enough so that endto-end delay remans below ; ths n turn leads to a de-, the crease n the number of streams. (Clearly, for + number of supported streams s zero.) In ths case, the number of streams supported ncreases sgnfcantly when the formaton tme s chosen optmally. Even though the analyss conducted above suggests a sgnfcant beneft n choosng the largest formaton tme possble, gven the end-to-end delay possble, the scenaro consdered s nevertheless unrealstc, as t assumes that the propagaton delay of all streams on the lnk s neglgble. Hereafter, we consder a more realstc scenaro n whch the propagaton delays correspondng to the streams on the lnk are random. As shown n Table II, we consder varous propagaton delay dstrbutons, dependng on the geographcal locaton of the lnk of nterest. If the lnk belongs to a local area network, then most of the calls have a relatvely low propagaton delay. On the other end of the spectrum, all connectons belongng to a trans-oceanc lnk have a large propagaton delay. (The dstrbuton of propagaton delay for streams flowng n a wde area network les somewhere between these two extremes.) In each of these cases, we assume that the formaton tme used for the packetzaton of each stream s talored to the propagaton and queung delay t ncurs. In order to capture the beneft obtaned from dynamcally choosng the formaton tme of a per-connecton bass, we plot n Fgure 12 the number of streams supported as a functon of the maxmum allowable formaton tme. Clearly, as the proporton of streams havng a large propagaton delay ncreases, the beneft of talorng the choce of formaton tme for each connecton decreases sgnfcantly: the plots n Fgure 12 show that per-flow custom packetzaton allows the number of streams to ncrease from around 1000 to more than 2500 on T1 lnks n local areas; however, the number of streams supported ncreases to around 2000 n wde areas under the same condtons, whle trans-oceanc lnks do not beneft at all from custom packetzaton. Also, most of the gan s acheved by ncreasng the propagaton delay from 10 to 30 ms. One can argue that an ncrease n delay by 20 ms s acceptable, gven the sgnfcant beneft obtaned from the ncrease. As shown n Fgure 13 (n whch we Fg. 12. Maxmum number of streams versus the maxmum formaton tme used by the encoder for T1 and T3 lnks (10 ms mnmum formaton tme allowed). Fg. 13. Maxmum number of streams versus the maxmum formaton tme used by the encoder for T1 and T3 (30 ms mnmum formaton tme allowed). plot the same set of graphs as n Fgure 12, but wth restrctng the mnmum formaton tme to 30 ms), the beneft from custom packetzaton beyond ms appears low: n WANs, bandwdth s usually largely avalable and propagaton delay s large; therefore, ths case s the least nterestng, snce the ncentve for bandwdth reducton s low, and the actual potental for bandwdth reducton s low too. On the other hand, the ncentve for bandwdth reducton s the largest n the case of trans-oceanc lnks, where the bandwdth s scare, and the cost of nstallaton hgh; unfortunately, as seen n Fgure 13, the large propagaton delays on the lnks dctate the choce of low formaton tmes. Fnally, the potental for bandwdth reducton s largest n local networks; even when restrctng the mnmum formaton tme to 30 ms, the ncrease n number of streams s stll sgnfcant (from around 2200 to more than 3000 on T3 lnks, around 75 to more than 90 on T1 lnks) when formaton tme s chosen on a perconnecton bass. Stll, one could argue that wth the advance n fber optcs technology, avalable bandwdth n local networks wll also be large. However, n the stuaton n whch T1 and T3 lnks are used to aggregate voce traffc n local LANs, dynamc packetzaton on a per-connecton bass could allow the support of a factor of 20% and 36% more streams on the T1 and T3 lnks,

10 respectvely; conversely, for a gven number of streams, the gan can allow the use of fewer T3 lnks, yeldng a sgnfcant cost decrease. Fnally, note that when voce traffc s not alone on the lnk but shares t wth other traffc usng ether PQ or WRR, then the queung delay s sgnfcantly larger, reducng n turn the number of streams supported and the potental gan provded by per-flow custom packetzaton. In summary, usng+ ms leads to a neglgble waste n resources, and consttutes a good compromse n terms of delay and bandwdth utlzaton. Repeatng the same study for G.711 and G leads to the same observatons. The recommended formaton tme for these encoders s 10 ms (that s, the recommended packet sze s 80 bytes) and 30 ms (that s, the recommended packet sze s 20 bytes), respectvely. VII. CONCLUSION In ths paper, we have focused on networks where a separate queue for voce traffc s provded. In ths context, we have frst descrbed the partcularty of voce traffc as compared to other traffc on the network. We then assessed voce delay, and looked at the effect of network parameters on the delay ncurred by voce traffc; n partcular, we focused on the effect of the resdual transmsson tme of non-voce packets of voce delay, and showed that t consttutes the largest porton of voce delay n case PQ s used to schedule voce traffc. We also showed the mportance of bandwdth to reduce the delay percentle ncurred by voce n a network, and concluded that network delay becomes a neglgble porton of end-to-end delay n case avalable bandwdth exceeds 10 Mbt/s. We then compared dfferent schedulng schemes (that s, PQ, WRR and the provson of dfferent crcut for voce traffc), and showed that PQ leads to the best compromse between bandwdth utlzaton and delay mnmzaton, as long as the preempton of non-voce packets s mplemented on sub-10 Mbt/s lnks. Fnally, we studed the effect of packet sze on voce delay and bandwdth utlzaton, and showed that a packet formaton tme of 30 ms for G.729A and G (that s, a packet sze of 30 and 20 bytes, respectvely) and 10 ms for G.711 (that s, a packet sze of 80 bytes) consttute a good compromse between low delay and effcent network utlzaton. REFERENCES [1] D. Awduche, Y. Rekhter, J. Drake, and R. Coltun, Mult-Protocol Lambda Swtchng: Combnng MPLS Traffc Engneerng Control Wth Optcal Crossconnects, IETF workng draft, work n progress, Nov [2] Y. Bernet, J. Bnder, S. Blake, M. Carlson, B. Carpenter, S. Keshav, E. Daves, B. Ohman, D. Verma, Z. Wang, and W. Wess, A Framework for Dfferentated Servces, IETF workng draft <draft-etf-dffservframework-02.txt>, work n progress, Feb [3] J.-C. Bolot, End-to-End Packet Delay and Loss Behavor n the Internet, Proceedngs of SIGCOMM 93, Sept. 93. [4] C. Bormann, PPP n a Real-tme Orented HDLC-lke Framng, RFC 2687, Sept [5] P. T. Brady, A Statstcal Analyss of On-Off Patterns n 16 Conversatons, Bell Syst. Tech. Journal, Vol. 47, pp , Jan [6] K. Bullngtom and J. Fraser, Engneerng Aspects of TASI, Bell Syst. Tech. Journal, Vol. 38, March [7] A. Charny and J.-Y Le Boudec, Delay Bounds n a Network wth Aggregate Schedulng, EPFL-DSC Techncal Report DSC2000/022, Aprl [8] J. Dagle and J. Langford, Models for analyss of packet voce communcatons systems, IEEE Journal on Selected Areas n Communcatons, Vol. SAC-4, No. 6, pp , Sept [9] A. Eckberg, The sngle server queue wth perodc arrval process and determnstc servce tmes, IEEE Transactons on Communcatons, Vol. COM-27, No. 3, pp , March [10] Recommendaton G.711, Pulse Code Modulaton (PCM) of Voce Frequences, ITU, Nov [11] Recommendaton G.726, 40, 32, 24, 16 kbt/s Adaptve Dfferental Pulse Code Modulaton (ADPCM), ITU, Dec [12] Recommendaton G.723.1, Speech Coders: Dual Rate Speech Coder for Multmeda Communcatons Transmttng at 5.3 and 6.3 kbt/s, ITU, March [13] Annex A to Recommendaton G.729, Codng of Speech at 8kbt/s usng Conjugate Structure Algebrac-Code-Excted Lnear-Predcton (CS- ACELP), Annex A: Reduced Complexty 8 kbt/s CS-ACELP Speech Codec, ITU, Nov [14] S. Golestan, A Self-Clocked Far Queueng Scheme for Broadband Applcatons, Proceedngs of IEEE INFOCOMM 94, pp , June [15] P. M. Gopal and B. Kadaba, A Smulaton Study of Network Delay for Packetzed Voce, n Proceedngs of GLOBECOM 86, Dec [16] P. Goyal, A. Greenberg, C. Kalmanek, W. Marshall, P. Mshra, D. Nortz, and K. Ramakrshnan, Integraton of Call Sgnalng and Resource Management for IP Telephony, IEEE Network, pp , May/June [17] J. Gruber and L. Strawczynsk, Subjectve Effects of Varable Delay n Speech Clppng n Dynamcally Managed Voce Systems, IEEE Transactons on Communcatons, Vol. COM-33, No. 8, Aug [18] IEEE 802.1D, Standard for Local and Metropoltan Area Networks: Meda Access Control (MAC) Brdges, [19] V. Jacobson, K. Nchols, and K. Podur, Expedted Forwardng PHB, RFC 2598, June [20] V. Jacobson, K. Nchols, and K. Podur, The Vrtual Wre Per-Doman Behavor, Internet workng draft <draft-etf-dffserv-pdb-vw-00.txt>, work n progress, July [21] Y. C. Jenq, Approxmatons for Packetzed Voce Traffc n Statstcal Multplexer, n Proceedngs of INFOCOM 84, Aprl [22] M. Karam, F. Tobag, On Traffc Types and Servce Classes n the Internet, n Proceedngs of GLOBCOM 2000, Dec [23] M. Karol and M. Hluchy, Usng a Packet Swtch for Crcut-Swtched Traffc: A Queueng System wth Perodc Input Traffc, IEEE Transactons on Communcatons, Vol. 37, No. 6, pp , June [24] M. Katevens, S. Sdropoulos, and C. Courcoubets, Weghted Round- Robn Cell Multplexng n a General-Purpose ATM Swtch Chp, IEEE Journal on Selected Areas n Communcatons, Vol. 9, No. 8, pp , Oct [25] L. Klenrock, Queung Systems-Vol. II: Computer Applcatons, New York; Wley, [26] M. Mandjes, K. van der Wal, R. Kooj, and H. Bastaansen, n Proceedngs of the 7th IFIP ATM and IP Workshop, June [27] G. Mercankosk, The Vrtual Wre Per Doman Behavor - Analyss and Extensons, IETF workng draft <draft-mercankosk-dffserv-pdbvw-00.txt>, work n progress, July [28] S. Moon, J. Kurose, P. Skelly, and D. Towsley, Correlaton of Packet Delay and Loss n the Internet, Techncal Report 98-11, Department of Computer Scence, Unversty of Massachussetts, Amherst, Jan [29] G. Ramamurthy, B. Sengupta, Delay Analyss of a Packet Voce Multplexer by the Queue, IEEE Transactons on Communcatons, Vol. 39, No. 7, pp , July [30] M. Shreddhar and G. Varghese, Effcent Far Queueng Usng Defct Round Robn, Proceedngs of SIGCOMM 95, Sept [31] K. Srram and W. Whtt, Characterzng superposton arrval processes n packet multplexers for voce and data, IEEE Journal on Selected Areas n Communcatons, Vol. SAC-4, No. 6, pp , Sept [32] L. Zhang, Vrtual Clock: A New Traffc Control Algorthm for Packet Swtchng Networks, Proceedngs of ACM SIGCOMM 90, Sept