Traffic Control in Atm

8/8/2019 Traffic Control in Atm

1/8

Traffic Control in High-speed ATM NetworksPeifang Zhou and Oliver W . W . YangSchool of Information Technology and EngineeringUniversity of Ottawa

Ottawa. OntarioCanada I < lN 6 N 5E-mail: yang Qsite.uottawa.ca

AbstractThzs paper pre sents a frameuiork of traffic control in

hzgh-speed AT.11 ne tuo rk s on boih connection level andcell leuel. O n the con nectio n level, we consider traf-fic classification, bandulzdth allocataon, call admzssioncontrol, and billing met hod . O n the cell leuel. we ez am -ine queueing archateciure. f l o w control, and schedulzng.W e also capture the interdependencaes am ong varzoustraffic control aspects t o ensure a proper network opera-t ion . The key contr ibut ion of this paper is an integratedmethodology io handle bu rsty data traffic. Based on theper-\% queueing archzte cture, we introduc e I ) a szm-ple banduidth allocation mechanism which requzres nocomplex computa t ion for call admzssion control, .2 animproved credit-basedflow contro l scheme which ensureslossless and congestzon-free cell iransport, 3) an inno-uatrve scheduling algorzthm wzth throughput guarantee.an d 4) an easy bill ing method which imposes tariff onbursty data flows without real-time measurem ent andprocessing.Keywords: AT11 netwo rking, t raf fic cont rol , qua l i ty ofservices i ssues , dat a t raff ic engineer ing, network cont ro lm a n a g e m e n t .

1 IntroductionA T N ne t works supp l y swi t ch ing an d t r anspor t i n fr a s-t r u c t u r e i n an i n t egra t ed f a sh i on t o a wi de mul t i t udeof b roadband mul t i -medi a se rv ices . To provide sucha wide spect rum of services wi th acceptable grades ofqual i ty of service ( Q o S ) , t is necessary to design flowregu l a t ion and r e source man ageme nt s chemes and algo-r i t hms . \ Y it h com muni ca t i on bandw i d t h we ll i n t o t heGb / s r ange , t he r a t i o be t ween propaga t i on de lay an dcel l t ransmiss ion t ime becomes very large, thereby ren-der ing ineffect ive many of the t raf f ic cont rol schemesthat r el y on t he f eedback f rom t h e ne t work t o r egu l a t et raf f ic f low. I t becomes obvious tha t new approachesneed t o he deve l oped in t h e a r ea of t raff ic con t rol .

M any broadband se rv i ces demand t he under l y i nghigh-speed n etworks to provide th em wi th t .heir char-acter i s t ical ly required levels of perform ance. Key Qo Sparamet e r s i nc l ude de l ay , t h roughput , ce l l loss r a t i o .e t c . E nforc i ng Qo S parameters requi res a complex in-t .e rp lay am ong ca ll admi s s i on con t ro l , han dwi d t h a l loca -t i o n , flow con t rol , que ueing, l ink sch edul ing, pr ior i t ips .discard pol ic ies , e tc . Th e design of the ent i re ent l -to-end system and the interact ion of var ious componei i t sa r e o f t en more i m por t a n t t han t he op t i mi za t i on o f in tl i-vidual e : lements . A re la ted i ssue is t,he tratleoff betw eeneffic iency and im plem entat io n complexi ty. I t is of tenpreferable to have s imp le schemes th at are less ef fic ientbu t can be eas i l y i mpl ement ed a t G b / s speeds t ha n tohave schiemes which are efficient but complex. \Vithinthis cont ;ext, we present a f r amework of traffic con trol inhigh-speed ATM networks on both connect ion level an dcell level. O n t ,he connection level, we consider trafficclassi f icat ion, band width al loca t ion, cal l admiss ion con-trol, an d bi l ling method . On th e cell level , we exam-i ne queue i ng a r ch i t ec t u r e , flow cont rol , and schedul ing.Unl ike other papers which tend to focus on a par t i cu l a rtopic , th is paper considers th e t raff ic control problemfrom thle inter -operat ion point of view, and capturest h e in terdependencies amo ng v ar ious t raff ic cont rol as-pec t s t o ensure a proper ne t work opera t , i on . T h e em-phas i s i s on t he deve l opment o f a traffic control schemet h a t can b e easi ly i mpl ement ed a t G b / s speeds in everynode t h roug hout h i gh- speed n e t works . T he key con t r i-bu t i on o f t h i s pap er i s an integrated methodology t o h a n -dle burs ty d at a t raf f ic . Based on t he pe r -VC queue i ngarchi tecture , we int roduce

0 a s i mpl e bandwi d t h a l l oca t i on mechan i sm whi chrequi res no complex comp utat ion for cal l adm iss ioncont ro l ,

0 an improved credi t -based f low cont rol scheme toensure loss less and congest ion-f ree cel l t ranspor t ,0 an innovat.ive scheduling algorit .hin \ v i t . h throrig11-p u t g u a r a n t e e , a n d

0-8186-9014-3/98 $10.000 998 IEEE 183


2/8

0 an easy bi l l ing method which imposes tar i f f onbur s t y d a t a f lows wi t hou t r ea l - ti me measu rementand process ing.

T h e r e st o f t h i s paper is organ i zed as follows. Section2 presen t s o u r proposed f rai iiework of t raf fic co nt rol onboth conne ct ion and cel l levels . Sect ion 3 conduc t s pe r -fo rmance eva l ua t i on o f our p roposed schem es . S ec ti on4 concludes the paper .

2 A Framework of Traffic Con-trol

T he o bject ive of any t , raff ic cont rol scheme s i s to pro-vide a s ta ble . lossless, congestion-f ree envi ron men t tocarry t ,raffic while achieving as much mul t i p l ex i ng ga i nas poss ible. In this sect ion. we presents a f ramework oft raf f ic cont rol on both connect ion level and cel l level .On the connect ion level : we address the i ssues of t raf -f ic c lass i f icat ion. band wid th al loca t ion. call admiss ioncont ro l (C'AC),and b i l li ng me t hod . \ \'e t hen t r ans l a t et he r equ i r ement s on t he con nec t i on l eve l i n t o t hose onthe cel l level . an d we consider the qu eueing arch i tectur ewi thin sw i tches . flow cont rol between swi tches , an d l inkschedul i ng .2.1 Connection-level traffic controlI n t h i s subsec t ion . we exam i ne t he i s sues of traffic c las -s i f ica t i on , bandwi d t h a l l oca t i on , ca ll adm i s s i on con t ro l !and bi l l ing method on the connect ion level of t raf f iccon t ro l .2.1.1 Tiaffic classificationATM sup por t s tr anspor t o f mi xed t ra f fi c and t he r e a r eessent ia l ly two types t raf f ic wi thin an ATM ne t work .One t ype i s con t i nuous s t r eam t ra f fi c such as voice andvideo, which is sens i t ive t o de lay bu t t o l e r an t t o a cer-ta in level of cel l loss . Another type is b u r s t y d a t a t r a f-fic such as IP packets . which i s sensi t ive to cell loss b u tcan accommodate delay. \Fr i th a wel l - tuned f low con-t ro l mechan i sm [ l o ] , we can el im inate cell loss fo r da t atraffic. T here fore we can classify traffic based on ly ont he de l ay a t t r i bu t e . I n our p roposed f r amewo rk , we d i-vide t raf f ic into two type s: delay-sensi t ive t raff ic anddelay- inse nsi t ive t raf f ic .2.1.2 Bandwidth allocationFeedback cont rol i s n o t appropr iate for delay-sensi t ivet raf f ic , because cel l s are only meaningful to the des-t i na t i on when t hey a r r i ve i n t i me , and t hey a r e con i -pletely useless i f they a re out of specif ied delay boun d-

, hroughput Link Capacity -__Delay-insensitiveTraffic

Time

F i gure 1: Bandwidth shar ing between delay-sensi t ivean d delay-insensit ive traffic.

ary. Feedback cont rol wi ll inevi tably int roduce UII -necessary delays , which can not be tolerated by thedelay-sensi t .ive t raf fic . Ins tea d, we can s im ply el imi-na te an y feedback con t rol on delay-sensi tive t raff ic , givet h e m p r io r i ty t r e a t m e n t , a n d l et t h e m g o t h ro u g h t h et r a n s p o r t a t i o n p i p e w i th o u t a n y i m p e d i m e n t. T h i s ist he bes t a netwo rk ca n do for del iver ing delay-sensi t ivecel ls . H owever this ar ran gem ent requi res t ,hat resourceshave t o be r e se rved a l ong t .he pa t h f rom t h e source tot he des t i na t i on . I n o t he r words , bandwi d t h shou l d herese rved and a l l oca t ed by t he peak r a t e .

A l t hough we use peak - r a t e a l l oca t i on fo r de l ay-sensi t ive traf f ic , we can s t i l l achieve high ban dw idth ut i -l izat ion by recycling unused ban dw idths: we f i l l i n t hegap be t ween actual bandwidth of delay-sensi t ive t raf f ican d physical l ink cap aci ty w i th delay- insensit ive t raff ic .T h e key i dea i s t h a t a ny bandw i d t h unused by de l ay-sens i t ive connec t i ons i s momen t a r i l y made ava i l ab l e todelay- insensi tive o nes . Th e scenar io is depicted in Fig-u re 1. Delay- insensi t ive t raff ic tend t o be burs ty a ndthey h ave di f ferent QoS r equ i r ement s . T hey have a highpeak- t o - ave rage r a t i o , and t hey a r e ma i n l y conce rnedwi t h i n s t an t aneous b a n d w i d t h t h r o u g h p u t . W h e n aconnec t i on source t r ansm i t s a t peak r a t e whi ch r e su l t sin a bur s t a r r i v i ng at an i n t e rmedi a t e node / swi t ch , weneed to i nc rease t he ban dw i d t h a l l oca t ed to t he con-nec t i on as m u c h as possi b le t o ge t t he bur s t ou t o f t heswi tch quickly. In othe r words. we need to devise aschedul i ng scheme whi ch can ad j us t han dwi d t h accord-i ng t o t h e s t a t e of t h e c o n n e c ti o n . O n t h e o th e r h a n d ,we have to recognize that each l ink in a n e t w o r k h a sa l i mi t ed capac i t y and i t is sha red am ong many con-nect ions , therefore we have t o t r ea t connec t i ons i n af a i r and squ are way an d p roper l y a ll oca t e unused band -wi d t h fo r each i nd i v i dua l connec t i on .

\Ve propose that each delay- insensi t ive connect ion

184


3/8

speci fy or negot ia te wi th the under lying network an av -erage t r a n s p o r t r a t e or bandw i d t h i t requ i r es dur i ng t heconnec t ion s s e t up t i me . T h i s ave rage bandwi d t h wil lbe gua ran t eed by t he ne twork . Nodes / swi tches i ns i det he ne t work shou l d r e - a l l oca t e any unused bandwi d t han d therefore they shal l allow a higher t ransmiss ion ratewhen a bur s t a r r ives . I n o t he r words , ins t an t aneou st r ansmi s s i on r a t e can be h i gher t han t he ave rage band-width which is agreed upon dur i ng t he connec t i on ss e t u p t i m e . T h i s is one of the connect ion- level requi re-me nts which need to be implem ented on the cel l- levelin each AT11 switch. IVe shall present a cell schedulinga l gor i t hm to sa t is fy t he above r equ i rement i n t he n es tsect on.2.1.3 Call admission controlBased on our proposed t raf f ic c lass i f icat ion and band-width al locat ion discussed in the previous subsect ions ,cal l admiss ion cont rol i s fa i r ly s imple and s t ra ight for-ward . A call or connect ion which requi res bandwidthbw will be adm it ted if the fol lowing condi t ion is sa t i s -f ied a t eve ry node/ swit ch a l ong t he pa t h f rom source t odes t i na t i on :

e x i s t i n gwhere CL is the outgoing l ink capaci ty. In oth er words ,a cal l /connect ion wi ll be admit ted i f i t s required b and -wi d t h does not exceed what is lef t in terms of band-wi d t h capac i t y . For a delay-sensit ive con nect ion. bwrepresents the peak rate . For a delay-insensit ive connec-t i on , bw i s the average t ransni i ss ion rate . I f a connec-t i on i s adm i t t ed . t he ban dwi d t h r equ i r ement bw repre-s e n t s t h e c o m m i t m e n t m a d e b y a node / swi tch t o t r ans -mit cells at r a t e bill calculated over a long per iod of t i m e .For clelay-iiisensit ,ive connection s, th e actu al thr ou gh -p u t d e p e n d s on t he source r a t e and t h e l oad condi t iona t each node . As we shall s e e l a t e r , t h e t h r o u g h p u tcan go much h i gher t han bw in a l ight load condi t ionbecause of an i nnova ti ve schedu l ing a l gor i t hm i mpl e -mented in every node. From another point of view, bwis t he m i n i mum t r ansmi s si on r a t e gua ran t eed by a n o d eunder the heavy load condi t ion.2.1.4 Billing methodCur ren t I n t e rne t en i p l oys an all yo u can ea t p r i c i ngmodel . However, giving everyone unl imi ted access toa network wi th f ini te bandwidth is a recipe for con-gest ion, as epi tomized by t ,he wel l -known phenomen onof .IVorld \Vide iV ait . I t becomes i nev i t ab l e t ha twe have to met,er n e t w o r k bandwidth usage, prov id ingqua l i t y of service (QoS) and maki ng ne t work pe r fo r -mance p r (d i c t ab l e . Rea l - ti me met e r ing is preferable ,

b u t it doesnt , work in practice, especially in high-speednetworks . Th e reason is qui te s imple: real - t ime measur-i ng genera t e s a s t ron omi ca l amou nt o f da t a whi ch needto be storNed an d processed. Th e s torag e and process ingpower reqjui red a re jus t to o dem and ing to be ful f i l led.In au t h or s po i n t of view. real - t ime meter ing i s unnec-essary. We can s imply charge each connect ion basedon the rake it requested. For delay-sensit ive connec-t ions , pr ic ing wi l l be based o n the peak rate . Such con-nect ions pay a p r e m i u m for a h i g he r b a n d w i d t h t h a nwha t mi gh t be neces sa ry , bu t t hey o b t a i n a higher pri-o r i t y an d r e source guaran t ee t o ach i eve mi n i mum de l ay .For delay- insensi t ive connect ions , pr ic ing will be basedon th e average rate speci fied d ur ing connect ions setu pt i me . T h e ne twork wi ll gua ran t ee t he bandw i d t h re -quested in c ong e s f e d s i tu at io n, but wi l l provide a highert h roughput i f t he l oad is l ight . In other words, delay-insensi t ive connect ions can reap s ta t i s t. ical niul t, iples inggain by rseusing handwidth left by delay-sensit ive con-nections. If a delay-insensit ive connection wants to in -crease the t ransmiss ion r ate provided by the network. i tcan simplly order a higher average handwidth and payaccording;ly. In su mm ary , our pr ic ing model is pract ical ,because i t doesn t requi re any real - t ime measi i remmtand process ing.

2.2 Cell-level traffic controlIn this subsect ion. we examine the i ssues of queue i ng .f low con t rol , and schedu l ing on the cell-level traffic cowt i o l .2.2.1 Queueing architectureIn most proposed swi tch archi tectures in l i terat ,ure[ l . 91, -YIM swi tches were por t -or iented. The y werebased on first-in, f irst-out (FIFO) principle: cellsf rom var i ous v i r t ua l connec t ions (VCs) leaving an out -pu t por t were o rgan i zed i n t o a FIFO queue and p ro-cessed in the order in which they ar r ived. Th e func-tionali t ,y of trad it io na l ATAI switches was l imi ted torou t i ng ce ll s f rom i np u t po r t s t o ou t pu t por ts . As weall know. t he ope ra t i ons o f AT M ne t works a r e no t j us tl i m i te d t o t h e t r a n s p o r t of cel l s f rom sources to thei rdestinati tons. ATM ne t works have t o hand l e AT M t r a f -fic efficiently an d reliably. ATM swi tch is an essent ia le l ement of a n AThI network, and i t should be designedin l ine wi th the operat io ns of th e network. We arguet h a t ATM switches should be traffic-flow-oriented andswi tches should han dle each t raf fic flow separately. NewAT \I sw i tches th at a re currently being developed wi l lprovide per-VC queue ing [ I O ] , and therefore they canhandl e each t r a f f i c flow or coni iection separately.

T h e s h i ft from por t -or iented design to flow-oriented

185


4/8

des ign i n t he s t ru c t u r e o f ATlI swi tches i s s igni f icant ,and i t has p rofound i mpl i ca t i ons on t he ope ra t i ons o fATM n etworks . Per-VC queueing i s the corners ton eof our pro posed f ramework of traff ic cont rol on th e celllevel . T h e other two comp onents of cel l- level traf fic con-t r o l , floiv cont rol mechanism and l ink schedul ing algo-r i t h m , a r e b a s e d on the per- \ C queueing archi tecture .2.2.2 Flow controlSince we use peak rate a l locat ion for delay-sens i t ivet . raf f ic .we can eliminate flow control on delay-sensi t iveconnec t i ons . Cel l s which belong to those connect ionsreceive pr ior i ty t re atm en t . and they wi ll t raverse in-ter im nodes/swi tches ins ide the net ,work wi thout anyhurd l e . Flow control is however necessary for delay-insensi t ive t raf f ic . We use average band wi dth for delay-i nsens i ti ve connec t i ons. bu t t he i ns t an t aneo us t h rough-pu t can he much h i gher t han t he ave rage . T here foreflow control is essential for delay-insensit ive connectionst o e n s u r e a lossless, congest ion-free envir on me nt.T h e credi t -based f low cont rol for A B R service w a s re -j ec t ed by the ATAI Forum in 1994, because i t requi redper- IyC queueing in swi tches and was considered toocompl ex t o i mpl emen t . Recen t deve l opment i n swi t chdes ign mad e i t pos si b le t o i mpl ement pe r -V C queue i ngin AT11 swi tches [ lo] . Credi t-based f low cont rol repre-sen t s t , r emendous advan tages ove r it.s r a t e -based coun-t e rp a r t . and i n t ,h i s paper we a rgue t ha t c r ed i t -basedf low cont rol should be used in the AT11 \VANS.

Credi t -hased f low cont , rol is fas t and ef fect ive. Con-t ro l t heory d i c t a t e s t ha t any con t ro l loop m u s t o p e r a t efas ter than the device i t cont rols . This is impossible inAT11 \i--ASs. where t he round- t r i p t i m e fo r any con t ro ls ignal i s l imi ted by the speed of l ight ( this is physicsand round- t r i p de l ay cant be i mproved! ) , and t he o f -fered t raf f ic load can f luctuate f a s t e r t h a n t h e r o u n d -t r i p t i me . I n t he r a t e -based a l gor i t hms . R hl cells areused for t raf f ic nianagement and all R11 cel l s esper i -ence rou nd- t r i p de l ays . I n o t he r words . an A B R trafficsource relies on t he ne t work-wi de knowl edge ( r e t u rn edR h l cells) to acljust its send i ng r a t e . and con t ro l s ont raf f ic can t be enforced quickly and ef fect ively due tothe long round- t r ip delay. This is why hop-by-hop con-t rol i s prefer red over end- to-end cont rol . Al l trafficsources can r e spond prompt l y once c r ed i t s a r e ava i l -ab l e f rom t he r ece ive r o f t he ne s t hop . w i t hou t wa i t i ngfor any feedback cont rol s ignal f roni the dest inat ion att h e f a r e n d .Anot he r adv an t ag e o f us ing a credi t -base flow con t ro lis t he g ua ra n t ee o f zero cel l loss. T h i s is ex t r emel y i m-por t an t for loss-sensit ive data t.raffic. .A s we all knowtha t for da ta t raf f ic. t ,he loss of even one cell can t r igge ra re t ransmiss ion o f t housant ls of cel ls . C red i t s represent

t he am oun t o f buf fe r space ava i l ab le a t t he n es t -ho p r e-ceiver. If t he t r ans mi t t e r does no t exceed t h i s quan t i t yof da ta , there i s no r i sk of buffer overf low. Even underext reme o ver loa dsj queue leng thes ins ide swi tches can tg row beyond w ha t t he c r ed i t s a l lowed . I n con t r as t .r a t e -based app roaches can no t guarantee zero cel l loss.Under e s t r em e over l oads , queues can eas i ly g row l a rge ,resul t ing in buffer overf low a nd cell loss. Credi t -basedschemes addres s the ro ot cause of congest ion, which ist he une spec t ed a r r i va ls f rom var i ous sources , and de -m a n d t h a t no cell can be t ransm i t ted to the receiverwi t hou t i t s pe rm i s s ion . As a result of this kind of strictres t r ic t ion. the feedback cont rol loop i s inherent ly stablean d no cell will ever be lost .

In our hop-by-hop credi t -based f low cont rol scheme.there are two credi t types . l ink credi t and VC c red i t .between two adjacent swi tches . Link credi t representst h e m a s i m u m n u m b e r of cel l s the upst ream swi tch cant r ansmi t (or t he max i mu m number o f cel ls t he do i vn-s t re am sn*i tch can receiv e) . wi thout overf lowing thebuffer in the downst ream swi tch. VC credi t is set on aper VC bas i s . and i t con t ro l s t he m as i m uni number o fcel ls t he ups t r eam swi t ch c a n send on t he VC (o r t h emas i n i um number o f ce l l s t he downs t r eam swi t ch canreceive on t h e V C ) . I n s u m m a r y , th e r e is one l ink credi ta n d m a n y VC credi t s between two adjacent swi tches .W i t h i n l ink c r ed i t a nd VC c red i t l i n i i t s, t he ups t r eamcan t r an smi t ce ll s on any VC.

I t i s wor thwh i le to poin t ou t our pro posal i s dif ferent .f rom t h e o r i g i na l c red i t -based flow cont rol scheme andcred i t upd a t e p ro t oco l (CUP) [6]. In [6], flow con t ro lwas receiver-or iented, credi t s main tained by th e senderwere updated by the receiver on a per -VC bas i s , andthere was no l ink credi t . However, our flow controli s both receiver- and sender-or iented: receiver cont rolssender s c r ed i t l i mi t s ( bo t h l ink c r ed i t an d VC c red i t s ) ,bu t s ender can se l ec t any VC to t r a n s m i t w i t h in t h el imi ts .

W e se t l i mi t s on VC c red i t s t o addres s t he i s sue o ffairness. so t h a t no s i ng l e VC or a group of VCs wi l ldomi na t e . T h e sum of i nd i v idua l VC c red i t s can bemuch grea t e r t ha n l ink c r ed i t . T h i s g ives t he up s t r eamswitch f lexibi l i ty as to which VC to dispatch cells. I tal lows a bur s t o f ce ll s t o b l as t t h rough t he swi t chif t raf f ic load i s mo derate . Th is i s very desi rable forbur s t y da t a t r a ff ic .

I n summ ary , we use peak r a t e bandwi d t h a l l oca ti onan d no f low cont rol for delay-sensit ive t raf f ic , an d h op-by-hop credi t -based f low cont rol for delay- insensi t ivet raf f ic . Delay-sensi t ive t raff ic are t ransp or ted wi th a nybarr ier , whi le delay- insen si t ive cel ls wi ll not b e a dm it -te d to the network af ter credi t s are exhaust .ed becauseof congest ions wi thin t .he network. In essence. all t raf -fic flow are wel l regulated . e i ther by peak-rat.e con-

186


5/8

s t r a i n t , or by avai lable credi t s . In this fashion, we getrid of th e root cause of congestion, which is t h e u n -predicted arrivals of traffic. The refore ATM networksoperate in a s table , congest ion-f ree, and loss less ( fordelay-insensit ive traffic) environment.2.2.3 S liedulingPer-VC qu eueing archi tecture faci l i ta tes schedul ing ofcel l t ransmiss ion, because scheduler can obtain queueinformat ion on a per-VC basis . I t is much eas ier for th escheduler to make decis ions to provide QoS g u a r a n t e efor each and eve ry VC connec t i on . I n t h i s subsec t i on ,we propose a schedul ing algo r i thm for per-VC queu eingATAI switches. which allows a much h i gher i ns t an t a -neous t h roughp ut t ha n t he ave rage . T h i s is one o f t herequi rements f rom th e connect ion- level traf f ic cont rolt h a t n e ed s to be implemented on th e cel l- level.

T h e opera t i on o f each ou t go i ng l ink is t ime-s lot ted.Each t ime s lot corresponds to the t ransmiss ion t ime ofone AT11 cell . T her e i s a l ink scheduler for every out -going l ink. Link scheduler selects cells for transmissiono n a f ram e basis ( logical ly) wi th iV s lots in a f r a m e . Fora connection i wi t h an ave rage bandw i d t h r equ i rementof bm , , t he s l o t r a t i on r, in a f r a m e is

b tu .r -2 S, i E { D S ,D I } .- CLwhere D S i s the set of delay-sensi tive connect ions a ndD I is the set of delay- insensit ive connect ions .A t t he beg i nn ing o f a f r ame , t he schedu l e r f o r an ou t -put l ink inspects all VC queues associated wi th th e l ink,an d schedules t ransmiss ion of cells in each VC q u e u e u pt o t h e VCs ra t ion . W e impose credi t -based flow con t ro lfor delay-insensit ive connections. Therefore cell trans-mission i s subje ct to th e avai labi l i ty of credi t . Th iscondi t ion is i mpl ied t h rougho ut t h i s subsec ti on a nd wewont repeat i t f rom now on.

Since we use peak-rate allocation for delay-sensit iveconnections, cells in delay-sensit ive VC queues neverexceed thei r ra t ion s . Let ni be th e number of cell s in th eVC queue for connect ion i a t t he beg i nn i ng o f a f r a m e ,D S represent the se t of delay-sensi tive connect ions , wehave

n j 5 rj , j E D S . ( 3 )For delay- insensi tive connect ions , th e nu mb er of cel l s

at the beginning of a f r ame can be l a rge r t han t he i rrations. \Ve divide delay-insensit ive connection in to twosets : DII a n d D I O . D I U designates the s et of delay-insensi t ive connect ions for which the numbers of cel l sa t t he beg i nn i ng o f a f ram e are under thei r ra t ions , i .e ..

D I U = {delay- insensi t ive connect ion k. n k 5 Q } .( 4 )

D I O represents t ,he set of delay-insensit ive connec-t ions for which th e num ber s of cell s a t the beginning ofa f rame are over thei r ra t ions , i .e . ,

D I O =: {delay- insensi t ive connect ion 1 , nl > r , } . ( 5 )From the the schedul ing point of view, some s lots ina n iV-slot f rame may be un-al located, and not a l l slotsa l l oca t ed to connect ions in D S a n d D I U are ut i l ized.We can calculate ext ra slots unused by connect ions int h e s e t s D S a n d DIG as follows:es = ( N - ri) +x(rj, n j , )+ x ( r . i . ,- n k , ) ,

i E {DS .D I } , ji E D S , hi E DII . ( 6 )1 J t k ,

To achieve mul t iplexing gain a mo ng connrct ions . weneed t .o re-allocate tho se ex t ra s lo ts lef t by connect ionsi n t h e s e t D S a n d D I l T . an d use t hem t o d i spa tch cellswhich belong t o connect ions in D I O . In the mean t ime.connect ions in D I P shou ld be credi ted for f reeing upslots which al lows the scheduler to t ransmi t cel l s foro t he r connec t ions . T h ey should be compensated by be-ing allowed to t ran sm i t m ore cel ls tha n rat ions la ter 011.D at a t raf fic are bursty. i Vhen a burs t ar r ives . a connec-t i on may experi ence t he t r ans i t i on f rom t he se t D I I - toD I O , ar id i t i s fa i r for the connect ion to c la im prrvi -ously unused rat ions to t ransmi t more cel l s a n d get thcburs t quickly out of the IC queue. In other words. t h escheduler should schedule more cel l t ransmiss ions forconnect ions which have n t used up thei r ra t ions . Th isshould be viewed as pa r t of scheduler s cornmitnient ofproviding average bandwidth to every delay- insensi t veconnect ion. For this purpose, we int roduce a gmnt vari-able . gi for connect ion i . to keep t rack of unused rat iono n a per-VC basis . T he value of gi is non-negat,ive. Itis init ial ized as 0 (zero) . I t i s incremented when fewercells thain co nnec tion i s r a t i on a r e s chedul ed a t t h e be -g i nn i ng of a f r ame , and i t i s dec rement ed when morecell s (u p to g k ) t ha n connec t ion i s ra t ion are scheduledd u r i n g a f rame.O u r proposed scheduler calculates ext ra s lots avai l -able in each f ram e, an d selects connect ions which haveg r u n t s a.vailable for t ran sm i t t ing more cel ls tha n thei rrat ion s . Th e selection i s based on the grant var iable .T h e l a rge r va l ue of g r a n t , th e higher pr ior ity a connec-t ion has . . If g r a n t s are used up for all connect ions butex t r a s l o t s a r e s t il l no t exhaus t ed . t he schedu le r w i llselect connect ions in t ,he set D IO for cell transmission.Based on ou r bil l ing policy t ,hat end users pay m ore forh i gher bandwi d t h . t h e schedu le r will di s t r i bu t e ex t r as l o t s based on t he b andw i d t h r equ i r ement. T h e h i gherbandwidth a connection requires. t h e higher priorityt he connect ion has. Here we need to i i it roduce anoth erab so rp t ion var iable to keep t , rack of cxt , ra bandw idt h

187


6/8

c o n s u m p t i o n on a per -VC bas is . T h e absorpt ion vari-a b l e , a i for connection i , is non-nega t i ve . I t s t a r t s a t 0(ze ro) ? an d i t i s increment .ed when a connect ion takesadv an t a ge o f ex t r a s l o t s ava i l ab l e i n a f r a m e t o t r a n s m i tmo re cel ls . T he smal ler value of a b so rp t io n , t he h i gherpr ior i ty a connection has i n consumi ng an y ex t r a s l o ts .To m aintain fa i rness amon g connect ions which have re-qui red di f ferent bandw idths a nd to ref lect tar i f f imposedon di f ferent bandwidths . an y ex t r a bandwi d t h l e f t byconnect ions in D S a n d D I U shou l d be d i s t r i bu t ed i npropor t i on t o bandwi d t h r equ i r ement s for connect ionsi n t he se t D I O . In other words , ext ra s lot usage needsto be nornial ized by th e bandw idth requi reme nt , . I f m ie x t r a slots are used by connect ion I . i t s ab sorp t i on va r i-ab l e a ( , houl d be upda t ed as follows:

( 7 )where Ii i s a posi t ive constant . , an d a1 i s t he upda t edva l ue o f (11.Note that calculat ion in the above schedul ing algo-r i t hm i s pe r fo rmed on a per connect ion basis . not on aper cel l bas is . This leads to a s igni f icant reduct ion incom put a t i ona l e f fo r t when t he a l go r i t hm is comparedwi th o the r ones [2. 7 , 81 which mark a t i me- s t amp onevery ar r iving cel l. Therefore ou r discipline is m u c hs i mpl e r t han v i r t ua l t i me- s t amp t ype of scheduling al-g o r i t h m s . T h i s m a k es o u r a l g o r i t h m a feasible candi-da t e fo r p r ac t i ca l i mpl ement a t i on i n pe r -VC queue i ngATM swi t ches .

3 Performanee AnalysisIn this sect ion. we conduc t pe r fo rmance ana l ys i s o f ourpropose d f ramew ork of t raff ic cont rol . We focus ont he fo l l owi ng mos t i mpor t an t QoS paramet e r s : delay,t h r o u g h p u t , an d cell loss r a t i o .

3.1 Delay and Delay Jitter BoundsIn a previous sect ion, we proposed a logical f rame s t ruc -t u r e o f .V cell slots for a l ink wi th capaci ty C L . T h ef r a m e d u r a t i o n is denot ed as T . S o w we de r ive t hee n d - to - e n d d e la y b o u n d , a s s u m i n g t h a t t h e s a m e N - s l o tf r a m e s t r u c t u r e a n d l in k c a p a c i t y CL are used wi thint h e ne t work . T h e end- t,o -end de l ay is defined t ,o be th et i m e e l apsed be t ween t h e a r r i va l of a cell a t the f i rs t -hopaccess swi tch and t ,he receipt of the cell a t t .he las t -hopaccess swi tch.

F i r s t we ob t a i n t he de l ay hound a t the f i rs t -hop ac-cess swi tch. LVe a ss um e t h a t h e r e a r e k connec t i ons

w h o s e p e a k r a t e s u m m a t i o n is l es s t h a n C L I N , i. e. ,k -LLC p r i I7:

A .i = lwhere p r i i s t he peak r a t e fo r connec t i on i . Becauset h e r a t e s u m m a t i o n of t h e a b o v e k connec t i ons is n otgrea t e r t h an t h e bandwi d t h a l l oca ted t o one s l o t in anN-s l o t f r am e s t ruc t u r e , we can a r r ange ce l ls f rom t hesek c o n n e ct i on s t o s h a r e t h e s a m e slot posi t ion in consec-u t i ve f r ames .

T h e r e a r e t w o c o m p o n e n ts fo r th e delay exper iencedby a cell ar r iving a t the f i rs t -hop access switch: thew a i t in g t i m e t o b e t r a n s m i t t e d a n d t h e s e rv ic e t i m e f o rt he t r ansmi s s ion of the cel l. We f i rs t calculate the wai t -ing t ime encoun tered by th e cell . In th e wors t case, onecell f rom eac h of the k connec t i ons a r ri ves a t t he s am et i m e , a n d t h e y all mi ss t he s t a r t o f a f r a m e . T h e y h a v et o w a i t for t h e s t a r t of the next f r ame . \.Ve niark th e cel lf rom one connec t i on as * t agged . T h i s t agged ce llh a s to wait for the finish of transmission of cells fromo t h e r ( k - 1) conne ct ions , and therefore th e . taggedcell exper ien ces the longest wa i t ing t ime, which i s

d, 5 T+ ( k- 1)T T = ( k+ 1)T. ( 9 )T h e first term in the above equ at ion represents the up-pe r bo und of wa i ti ng t ime for t h e s t a r t of a new f r ame .t he second t e rm i s t he t i me e lapsed t o t r ansmi t ( k- 1 )ce ll s i n t he sa me s l o t pos i t ion . and t he l a s t t e rm is t h eu p p e r b o u n d b e tw e e n t h e s t a r t o f f r a m e a n d t h e s t a r tof t r ansmi s s i on o f t he t agged ce ll .

Now we ca lculate the second com ponen t of the c lelayat the f i rs t -hop access swi tch. which is the cel l t rans-miss ion t ime. Let 1 denote the cell length. Th e cel lt r ansmi s s i on t i me m is 1

C Lm = -.Combi n i ng t he above t wo t e rms , we can expres s t hede l ay d l , at th e f i rs t-hop access swi tch as

S i nce ce l l s f r om t he above 12 connec t i ons sha re t hesame s l o t pos i t i on , t he superpos i t i on of these k con-nect ions wi l l be t reated as a s ingle connect ion comingou t of t h e acces s swi t ch an d going int o core of th e ne t -work . Reus i ng eqn . (9), we have delays int roduced att he core swi t ches w i th a m a x i m u m o f 2T each. If therea r e h hops be t ween t he f i r st - hop acces s swi t ch and t hel as t -hop acces s swi t ch , t he ma xi m um de l ay i n t roducedby swi tches oth er th an the f i rs t one wi ll be h x 2T. T h edelay i t i t roduced by all swit.ches is given by

1CLd,, = di , i-x 2T I k + l )T+- h x 2T. (12 )

188


7/8

Addi ng t he t he p ropaga t i on de l ay p r o p t o t h e a b o v ed,,, we wi l l a t ta in the end- to-end delay d ,

1d = p r ~ p + d , , 5 p r o p + ( I :+ 1 ) T + c'+ h x 2T . (13)LBased on t he above ana l ys i s and de r i va t i on , we can eas-i ly ob t a i n t h e upper bound and l ower bound of t h e e n d -to-end delay d :

Id,,,ax 5 p r o p + ( I :+ 1 ) T +- h x 2T (14)CLT h e d e la y j i t t e r dj is defined as the di f fcrence betweend,,,, an d d,,,,. T h e upper bou nd fo r de lay j i t t e r canbe t ler i \et l f r o m e q n . ( 1 4 ) a n d e q n . (1.5)as follows:

I f the ra te of a connect ion is so l a rge t ha t at least twocell s have to be forwarded in a f ranie . we can decom poset he connec t i on i n t o sub-connec t i ons , s u c h t h a t we canuse t he above de r i va t i on to obt a i n s i mi l a r r e su l t s.

3.2 ThroughputO u r per - \ 'C queue i ng schedul er (P V Q S ) has a t t r ac t i v eprope r t ies th at can be expressed by th e following lem-inas a n d t h e o r e m .Lemma 1: P\'QS prov i des bandwi d t h guaran t ee forevery connect ion.Proof: For an out go i ng l i nk wi t h capac i t y C'L a n d N( logical ) s lots in a f r ame . we al locate slot r a t i on T ; forconnect ion i according to eqn. ( 2 ) :

b 111.r . - 2 x . . ie { D S . D I )- C ' LT h e ba i i dwi d t h a l l oca t ed by P VQ S t o connec t i on i is

Loinilia 2: P\'QS is f a i r in t h e sense t ha t i t a l l oca t esany es t ra hai lc lwidth in propo r t ion to requested b and -w i t l t l i s .

Proof: 'LVe use t h e absorpt ion var i ab l e t o keep t r ackof e x t r a s l o t u s a g e. If t he r e a r e any ex t r a s l o t s l e f tby connect ions in D S a n d D I U , P V Q S s o r t s o u t c o n-nec t i ons i n t he se t D I O accord i ng t o t he a b so rp t io nvar i ab l e . Co nnec t i ons w i t h l e s s absorpt ion values havea h i gher p r i o r i ty t o use ex t r a s l o t s t o t r ansmi t morecel ls . A f ter cel l t ransmiss ion, a connection's absorpt ionvar iable is i nc r eased an d t he connect ion has l e ss chancet o use any more s l o t s . I n t h e l ong run , t he d i f fe r encesa m o n g a b so rp t io n var iab l es a r e much smal l e r t han ab-sorp t ion var iables themselves . Let I a n d J r epresen te x t r a sloi s consumed by connect ions i a n d j d u r i n g t h ei n t e rva l ( 0 ,x). gnor ing the smal l di f ference betweena b so rp t io n var i ab l es al a n d a,, we have

Q.E.D.

Theorern: For delay-insensit ive connection i , t h e es -tra bandwi d t h r ece i ved beyond t he requested averagebandwi d1 h bw, is

Proof: From Lemma 2. any r emai n i ng bandw i d t h ca -paci ty, which i s CL- Cl w, -cl u.,, i E D I . j ED S . will be dis t r ibuted in propo r t ion to requested av-e r age r a t e s of the delay- insensi t ive connect ions . Fordelay- insensi t ive connect ion i which has requested ar a t e bw,. i t s por t ion in total requested average ban d-w i d t h is bw,/ E, b w, , therefore connection i can benef i tf rom t h e fo l lowi ng ex t r a band wi d t h beyond bw,:

bw ;extra b a n d w i d t h =- (CL- bwj - b w j ) ,Xi bwi i ji E D I , j E D S. Q.E.D.

3.3 Cell loss ratio\Ve consider t h e buffer requirement for delay-sensit iveand dela ,y- insensi t ive connect ions in the core swi tches .For delay-se nsi t ive conn ect ions , we use peak-ra te allo-ca t i on an d r e se rve enough slots accord i ng ly . T h e mas i -m u m n u m b er of delay-sensit ive cell arrivals in an .V-slotf r a m e is ,V . Therefore the buffer ing requi rement :\IDSfo r delay-sen si t ive connect ioi is is N cells, i .e. ,

. U D ~ ,V, for delay-sensit ive con nect ions . ( I d )

189


8/8

Now n e consider t he case for delay- insensi tive f lows.In our f low cont rol proposal , l ink credi t represents themax iniuin buffer space avai lable in the receiver (dow n-s t r e a m s w i t c h ). As long as t he sender (upper s t r eamswi t ch) does no t s end more ce l l s t han wha t t he l i nkcredi t allows. no buffer overflow will occur in the re-ceiver.

HoweLer, receiver has to t ake i n t o account t he l i nkdelay before sender receives any l ink credi t update . Int he wor s t case , t he max i mum nu mber o f cel ls A U o ~h a tcan be t ran sm i t ted by the sender before i t receives feed-back f rom the receiver i s

where CL i s the l ink capaci ty and RTT i s the round-t r i p t i me be t ween a sender and a receiver. T here fore ,t he m i n i m um buff er s i ze r equ i red a t t h e dow ns t r eamswitch to prevent buffer overflow is given by

As long as t he downs t r eam swi t ch ma i n t a i ns a buffersize of at least B,,,i,, there i s no cell loss.

In suniniary. our proposed f ramework of t raf f ic con-t ro l i mposes an uppe r bound on t h e e n d - b e n d d el ay fo rdelay-sensi tive co nnect ions . I t also prov i des t h roughputguara ntee for delay- insensi tive con nect ions . and ensureszero cell loss.

4 ConclusionIn this p ap er , we int, roduced a framework of traffic con-t rol in high-speed A T lI networks on bot h connec t i onlevel and cell level. On the connection level, we consid-ered t raff ic c lassi f icat ion, bandw idth al locat ion, cal l ad-miss ion cont . rol. an d bi l ling metho d. On the cell level,we exami ned queue i ng a r ch i t ec t u r e , flow con t ro l , andschedu l ing. Th is pap er considered the t raf fic cont rolp rob l em f rom t he i n t e r -opera t i on po i n t o f v i ew , andcaptu red the int .erdependencies am on g various t raf f iccon t ro l a spec t s t o ensure a proper ne t work opera t i on .T he key con t , r i bu t i on of t h i s paper was an i n i e g r a t e dm e t h o d o l o g y t,o handl e bur s t y da t a tr a ff ic . Based ont he pe r -VC queue i ng a r ch i t ec t u r e , we i n t roduced 1) as imp le bandwicl th a l locat ion mechanism which requi redno compl ex comput a t i on for cal l admiss ion cont rol , 2)an im proved cred i t -based f low cont rol schem e which en-sured loss less and congest ion-f ree cel l t ranspor t , 3 ) ani nnova t i ve schedu l i ng a l gor i t hm wi t h t h roug hput gua r -a n t e e . a n d 4 ) an easy bi l l ing method w h i c h i mposedtariff on Itursty (lata flows wit ,liout real-t inie m easu re-

ment and p roces si ng .

Ack ow1 dgmentT h i s p a p e r is par t i a l l y suppor t ed by an NSERC Oper -a t i n g G r a n t u n d e r c o n t r a c t #OGP0042878.

References[l] R. A w d e h and H. S fouf t ah . S urvey of AT31 swi tch

a rch i t ec t u r es , C o m p u t e r lVetworks ar id ISD!V Sys-t e m s , vol. 27, pp. 1567-1613. Nov. 1995.

[2] .J. Benne t t and H . Z hang , W F 2 Q : worst-casefai r weight fa i r queu eing, . in Proc. I E E E I S F O -COJf96, San Francisco, C A , Mar. 1996, pp 120-128.[3] F . B o n o m i a n d K . IV . Fendick, T he rate-basecl flow

control f ramework for the avai lable hi t ra te AT11service. I E E E 1VetworX: JIaga:ine, vol. I). pp . 2%-39 . 1 f a r . / Apr . 1995 .

[4] AT11 Forum . AT31 Traff ic h Ianagenien t Speci fica-t ion. Vers ion 4.0. Apr . 1996.

[5] L. M ei nrock , T h e l a t ency / bandwi d t h t radeoff i ng i gab i t ne t works , I E E E Comm. :2iiaga:zne, vol. 30 ,pp . 36-40 , Apr . 1992 .

[6] H. T. K u n g a n d R. Morris, ;Credit-based flow con-t rol for ATM networks , I E E E N e t w o r k M a g a : i n e ,vol . 9 . pp. 40-48, M ar . / Apr . 1995 .

[7] A. P a r e k h a n d R . Gal l age r , A generalized proces-sor sha r i ng approach to f low cont rol in integ ratedservices networks- the s ingle node case, in Proc.I E E E IiVFOCOAf 92. Florence, I ta ly, May 1992 , pp.9 1 .5924 .

[SI L . Z hang , Vi r t ua lC l ock : a new traffic control algo-r i th m for packet swi tching netwo rks , in Proc. AC,\lSIGCrOMM90, P hi l ade l ph i a , PA, S ept . 1990 , pp .19-29.

[9] P. Z h o u a n d 0. Y a n g , A new design of cent ral -queueing ATM swi tches , in Proc. IEEE G L O B E -COll197, N o v . 1997, P hoeni x , AZ, pp . 541-545.

[lo] P. Z hou and 0 . Yang, Design of per-VC queue-ing XTM sw i tches , in Proc. I E E E ICC98, A t l a n t a ,G X , J u n . 1 9 98 , p p . 304-308.

f l l ] P. Zhou and 0 .Y a n g , A f ramework off low co nt rolin high-speed ATSI networks, To appear i n Proc.I E E E dlILCOAI98. Bedford , M A , Oct . 1998.

190

Traffic Control in Atm

Documents