Design Considerations for the Virtual Source/Virtual Destination (VS/VD) Feature in the ABR Service of ATM Networks

Design Considerations for the Virtual Source/Virtual Destination

(VS/VD) Feature in the ABR Service of ATM Networks

Shiv Kalyanaraman, Raj Jain, Jianping Jiang, Rohit Goyal, Sonia Fahmy

The Ohio State University

Department of CIS, Columbus, OH 43210-1277

Phone: +1 614 688 4482. Fax: +1 614 292 2911

Email: fshivkuma, jain, goyal, [email protected]

Pradeep Samudra

Director, System Engineering,

Broadband Network Lab, Samsung Telecom America, Inc.

1130 E Arapaho, Richardson, TX 75081.

Phone: (972) 761-7865, email: [email protected]

Keywords: Asynchronous Transfer Mode (ATM), Available Bit Rate (ABR),

Virtual Source/Virtual Destination (VS/VD), Tra�c Management, Congestion Control.

Abstract

The Available Bit Rate (ABR) service in ATM networks has been speci�ed to allow fair and e�cient

support of data applications over ATM utilizing capacity left over after servicing higher priority

classes. One of the architectural features in the ABR speci�cation [1] is the Virtual Source/Virtual

Destination (VS/VD) option. This option allows a switch to divide an end-to-end ABR connection

into separately controlled ABR segments by acting like a destination on one segment, and like a

source on the other. The coupling in the VS/VD switch between the two ABR control segments

is implementation speci�c. In this paper, we model a VS/VD ATM switch and study the issues in

designing coupling between ABR segments. We identify a number of implementation options for

the coupling. A good choice signi�cantly improves the stability and transient performance of the

system and reduces the bu�er requirements at the switches.

1 Introduction

Asynchronous Transfer Mode (ATM) networks provide multiple classes of service tailored to support

data, voice, and video applications. Of these, the Available Bit Rate (ABR) and the Unspeci�ed Bit

Rate (UBR) service classes have been speci�cally developed to support data applications. Tra�c is

1

controlledintelligentlyinABRusingarate-based

closed-loop

end-to-endtra�

cmanagem

entfram

e-

work[1,2,

3].Thenetworkswitches

monitor

availablecapacityandgive

feedbackto

thesources

askingthem

tochange

theirtransm

ission

rates.

Several

switch

algorithmshavebeendeveloped

[4,5,6,7,8]to

calculatefeedbackintelligently.Theresourcemanagem

ent(RM)cells(whichcarry

feedbackfrom

theswitches)travelfrom

thesourceto

thedestinationandback.

Oneof

theoptionsof

theABR

fram

eworkis

theVirtual

Source/Virtual

Destination(V

S/V

D)

option.Thisoption

allowsaswitch

todividean

ABRconnection

into

separatelycontrolledABR

segm

ents.Ononesegm

ent,theswitch

behavesas

adestinationendsystem

,i.e.,itreceivesdataand

turnsaroundresourcemanagem

ent(RM)cells(whichcarryratefeedback)to

thesourceendsystem

.

Ontheother

segm

enttheswitch

behaves

asasourceendsystem

,i.e.,itcontrolsthetransm

ission

rate

ofeveryvirtual

circuit(VC)andschedulesthesendingof

dataandRM

cells.

Wecallsuch

aswitch

a\V

S/V

Dswitch".

Ine�ect,theend-to-endcontrol

isreplacedbysegm

ent-by-segment

controlas

show

nin

Figure

1.

Figure

1:End-to-EndControlvsVS/V

DControl

Oneadvantage

ofthesegm

ent-by-segmentcontrol

isthat

itisolates

di�erentnetworksfrom

each

other.Oneexam

pleisaproprietarynetworklike

fram

e-relayor

circuit-switched

networkbetween

twoABRsegm

ents,whichallowsend-to-endABRconnection

setupacrosstheproprietarynetwork

andforwardsATM

packets

betweentheABRsegm

ents

1.Another

exam

pleistheinterfacepoint

betweenasatellitenetworkandaLAN.Thegateway

switches

attheedge

ofasatellitenetwork

canimplementVS/V

Dto

isolatedow

nstream

workgroupswitches

from

thee�ectsofthelongdelay

satellitepaths(likelongqueues).

Asecondadvantage

ofsegm

ent-by-segmentcontrolisthat

thesegm

entshaveshorterfeedbackloops

whichcanpotentiallyimprove

perform

ance

becausefeedbackisgivenfasterto

thesourceswhenever

new

tra�

cburstsareseen.

1SignalingsupportforthispossibilityisyettobeconsideredbytheATM

Forum

2

The VS/VD option requires the implementation of per-VC queueing and scheduling at the switch. In

addition to per-VC queueing and scheduling, there is an incremental cost to enforce the (dynamically

changing) rates of VCs, and to implement the logic for the source and destination end system rules

as prescribed by the ATM Forum [1].

The goal of this study is �nd answers to the following questions:

� Do VS/VD switches really improve ABR performance?

� What changes to switch algorithms are required to operate in VS/VD environments?

� Are there any side-e�ects of having multiple control loops in series?

In this paper, we model and study VS/VD switches using the ERICA switch algorithm [8] to

calculate rate feedback. We describe our switch model and the use of the ERICA algorithm in

sections 2 and 3. The VS/VD design options are listed and evaluated in sections 4 and 5. The

results and future work are summarized in sections 7 and 8.

2 Switch Queue Structure

In this section, we �rst present a simple switch queue model for the non-VS/VD switch and later

extend it to a VS/VD switch by introducing per-VC queues. The ow of data, forward RM (FRM)

and backward RM (BRM) cells is also closely examined.

2.1 A Non-VS/VD Switch

A minimal non-VS/VD switch has a separate FIFO queue for each of the di�erent service classes

(ABR, UBR etc.). We refer to these queues as \per-class" queues. The ABR switch rate allocation

algorithm is implemented at every ABR class queue. This model of a non-VS/VD switch based

network with per-class queues is illustrated in Figure 2.

Besides the switch, the �gure shows a source end system, S, and a destination end system, D, each

having per-VC queues to control rates of individual VCs. For example, ABR VCs control their

Allowed Cell Rates (ACRs) based upon network feedback. We assume that the source/destination

per-VC queues feed into corresponding per-class queues (as shown in the �gure) which in turn feed

to the link. This assumption is not necessary in practice, but simpli�es the presentation of the

3

Figure

2:Per-classqueues

inanon-VSVDswitch

model.Thecontention

forlinkaccess

betweencellsfrom

di�erentper-classqueues

(attheswitch,

thesourceandthedestination)isresolved

through

appropriatescheduling.

2.2

AVS/VD

Switch

TheVS/V

Dswitch

implementsthesourceandthedestinationendsystem

functionalityinaddition

tothenormalswitch

functionality.Therefore,like

anysourceanddestinationend-system

,itrequires

per-VC

queues

tocontrol

theratesof

individual

VCs.

Theswitch

queuestructure

isnow

more

similar

tothesource/destinationstructure

wherewehaveper-VCqueues

feedinginto

theper-class

queues

beforeeach

link.Thisswitch

queuestructure

andaunidirectional

VCoperatingon

itis

show

nin

Figure

3.

TheVS/V

Dswitch

has

twoparts.Thepartknow

nas

theVirtualDestination(V

D)forwardsthe

datacellsfrom

the�rstsegm

ent(\previousloop")

totheper-VCqueueat

theVirtualSource(V

S)

ofthesecondsegm

ent(\nextloop").Theother

partor

theVirtualSource(ofthesecondsegm

ent)

sendsoutthedatacellsandgenerates

FRM

cellsas

specifed

inthesourceendsystem

rules.

Figure

3:Per-VCandper-classqueues

inaVSVDswitch

Theswitch

also

needsto

implementtheswitch

congestioncontrol

algorithm

andcalculate

the

allocationsforVCsdependingupon

itsbottleneckrate.Aquestion

whicharises

iswheretherate

4

calculationsaredoneandhow

thefeedbackisgivento

thesources.Wepostponethediscussionof

thisquestion

tolatersections.

2.3

AVS/VD

Switch

withUnidirectionalData

Flow

Theactionsof

theVS/V

Dswitch

upon

receivingRM

cellsareas

follow

s.TheVDof

theprevious

loop

turnsaroundFRM

cellsas

BRM

cellsto

theVSon

thesamesegm

ent(asspeci�ed

inthe

destinationendsystem

rules[2]).Additionally,when

theFRM

cellsareturned

around,theswitch

may

decreasethevalueoftheexplicitrate

(ER)�eldto

accountforthebottleneckrate

ofthenext

linkandtheERfrom

thesubsequentABRsegm

ents.

When

theVSat

thenextloop

receives

aBRM

cell,theACR

oftheper-VCqueueat

theVSis

updated

usingtheER�eldin

theBRM

(ERof

thesubsequentABRsegm

ents)as

speci�ed

inthe

sourceendsystem

rules[2]).Additionally,theERvalueofthesubsequentABRsegm

entsneedsto

bemadeknow

nto

theVD

ofthe�rstsegm

ent.

Oneway

ofdoingthisisfortheVD

ofthe�rst

segm

entto

use

theACRoftheVCin

theVSofthenextsegm

entwhileturningaroundFRM

cells.

Themodelcanbeextended

tomultipleunidirectionalVCsinastraightforwardway.Figure4show

s

twounidirectionalVCs,VC1andVC2,betweenthesamesourceSanddestinationDwhichgo

from

Link1to

Link2on

aVS/V

Dswitch.Observethat

thereisaseparateVSandVDcontrol

foreach

VC.Weom

itnon-ABRqueues

inthisandsubsequent�gures.

Figure

4:MultipleunidirectionalVCsin

aVSVDswitch

5

2.4

Bi-directionalData

Flow

Bi-directional ow

inaVS/V

Dswitch

(Figure

5)isagainasimpleextension

totheabovemodel.

Thedataon

thepreviousloop

VDisforwarded

tothenextloop

VS.FRMsareturned

aroundby

thepreviousloop

VDto

thepreviousloop

VS.BRMsareprocessed

bythenextloop

VSto

update

thecorrespondingACRs.

Figure

5:Multiplebi-directionalVCsin

aVSVDswitch

Wewilldiscuss

theratesandallocationsof

VC1only.VC1has

twoACRs:ACR

1in

thereverse

direction

onLink1andACR

2in

theforwarddirection

onLink2.

Henceforth,thesubscript1refers

tothe\previousloop"variablesandsubscript2to

the\nextloop"variablesof

VC1.

3BasicERICASwitchScheme

Weuse

theERICAalgorithm

[8]forcongestioncontrol

attheswitches.Wegive

abriefoverview

ofthealgorithm

inthissection.

ERICA�rstsetsatarget

rate

asfollow

s:

TargetRate=TargetUtilization

�LinkRate-VBRRate-CBRRate

Italso

measurestheinputrate

totheABRqueueandthenumber

ofactive

ABRsources.

Toachieve

fairness,theVC'sAllocation

(VA)has

acomponent:

VAfairness=TargetRate/Number

ofActiveVCs

Toachieve

e�ciency,theVC'sAllocation

(VA)has

acomponent:

VAe�

ciency

=VC'sCurrentCellRate/Overload,whereOverload=InputRate/TargetRate;

Finally,theVC'sallocation

onthislink(VAL)iscalculatedas:

6

VAL=MaxfVAe�

ciency,VAfairnessg=FunctionfInputRate,VC'scurrentrateg

Notethat

thefullERICA

algorithm

containsseveralenhancements

whichaccountforfairness,

queueingdelays,andwhichhandleshighlyvariantbursty

(ON-OFF)tra�

ce�

ciently.

Acomplete

description

ofthealgorithm

isprovided

inreference

[8].Wenow

describewheretheERICA

rate

calculationsaredonein

anon-VS/V

Dswitch

andin

aVS/V

Dswitch.

3.1

Rate

Calculationsin

anon-V

S/VD

Switch

Thenon-VS/V

Dswitch

calculatestherate

(VAL)forsources

when

theBRMsareprocessed

inthe

reversedirection

andentersitin

theBRM

�eldas

follow

s:

ERin

BRM

=MinfERin

BRM,VALg

Atthesourceendsystem

,theACRisupdated

as:

ACR=FunctionfER,VC'scurrentACRg

3.2

Rate

Calculationsin

aVS/VD

Switch

Figure6show

stheratecalculationsinaVS/V

Dswitch.Speci�cally,thesegm

entstartingat

Link2

(\nextloop")

returnsan

ERvalue,ER

2in

theBRM,andtheFRM

ofthe�rstsegm

ent(\previous

loop")

isturned

aroundwithan

ERvalueofER

1.TheERICA

algorithm

fortheportto

Link2

calculatesarate

(VAL2)as:VAL2=

FunctionfInputRate,

VC'sCurrentRateg.Therate

calculationsat

theVSandVDareas

follow

s:

Figure

6:Ratecalculationsin

VS/V

Dswitches

�Destinationalgorithm

forthepreviousloop:

ER

1=MinfER

1;VAL2;ACR

2g

7

�SourceAlgorithm

forthenextloop:

Optionally,ER

2=MinfER

2;VAL2g

ACR

2=FnfER

2;ACR

2g

Theunknow

nsin

theaboveequationsaretheinputrate

andtheVC'scurrentrate.Weshallseein

thenextsectionthat

thereareseveralwaysof

measuringVCratesandinputrates,combiningthe

feedbackfrom

thenextloop,andupdatingtheACRof

thenextloop.Notethat

though

di�erent

switches

may

implementdi�erentalgorithms,manymeasure

quantities

such

astheVC'scurrent

rate

andtheABRinputrate.

4VS/VDSwitchDesignOptions

Inthissection,weaim

atansweringthefollow

ingquestions:

�What

isaVC'scurrentrate?(4

options)

�What

istheinputrate?(2

options)

�Does

thecongestioncontrol

actionsat

alinka�ectthenextloop

orthepreviousloop?(3

options)

�When

istheVC'sallocation

atthelink(VAL)calculated?(3

options)

Wewillenumeratethe72

(=4�2�3�3)

option

combinationsandthen

studythisstatespace

forthebestcombination.

4.1

MeasuringtheVC'sCurrentRate

Therearefourmethodsto

measure

theVC'scurrentrate:

Figure

7:Fourmethodsto

measure

therate

ofaVCat

theVS/V

Dswitch

8

1.Therate

oftheVCisdeclaredbythesourceendsystem

ofthepreviousloop

intheCurrent

CellRate(CCR)�eldof

theFRM

cell(FRM1)

received

bytheVD.Thisdeclaredvaluecan

beusedas

theVC'srate.

2.TheVSto

thenextloop

declarestheCCR

valueof

theFRM

sent(FRM2)

tobeitsACR

(ACR

2).Thisdeclaredvaluecanbeusedas

theVC'srate.

3.Theactual

sourcerate

inthepreviousloopcanbemeasured.Thisrate

isequal

totheVC's

inputrate

totheper-VCqueue.

Thismeasuredsourcerate

canbeusedas

theVC'srate.

4.Theactualsourcerateinthenextloopcanbemeasuredas

theVC'sinputrateto

theper-class

queue(from

theper-VCqueue).Thismeasuredvaluecanbeusedas

theVC'srate.

Figure

7illustrateswhereeach

methodisapplied

(notethepositionof

thenumbersin

circles).

4.2

MeasuringtheInputRate

attheSwitch

Figure

8(notethepositionof

thenumbersin

circles)

show

stwomethodsof

estimatingtheinput

rate

foruse

intheswitch

algorithm

calculations.Thesetwomethodsare:

Figure

8:Twomethodsto

measure

theinputrate

attheVS/V

Dswitch

1.Theinputrate

isthesum

ofinputratesto

theper-VCABRqueues.

2.Theinputrate

istheaggregateinputrate

totheper-classABRqueue.

4.3

E�ectofLinkCongestionActionsonNeighboringLinks

Thelinkcongestioncontrolactionscana�ectneighboringlinks.Thefollow

ingactionsarepossible

inresponse

tothelinkcongestionof

Link2:

9

1.ChangeER

1.Thisa�ectstherateofthepreviouslooponly.Thechange

inrateisexperienced

only

afterafeedbackdelay

equalto

twicethepropogationdelay

oftheloop.

2.ChangeACR

2.Thisa�ectstherate

ofthenextlooponly.Thechange

inrate

isexperienced

instantaneously.

3.Change

ER

1andACR

2.Thisa�ects

both

thepreviousandthenextloop.Thenextloop

is

a�ectedinstantaneouslywhilethepreviousloop

isa�ectedafterafeedbackdelay

asin

the

�rstcase.

4.4

FrequencyofUpdatingtheAllocatedRate

TheERICA

algorithm

inanon-VS/V

Dswitch

calculatestheallocatedrate

when

aBRM

cellis

processed

inaswitch.How

ever,in

aVS/V

Dswitch,therearethreeoptionsas

show

nin

Figure

9:

Figure

9:Threemethodsto

updatetheallocatedrate

1.Calculateallocatedrate

onreceivingBRM2only.Storethevalueinatableandusethistable

valuewhen

anFRM

isturned

around.

2.Calculate

allocatedrate

only

when

FRM1is

turned

around.

3.Calculateallocatedrate

both

when

FRM1isturned

aroundaswellaswhen

BRM2isreceived.

Inthenextsection,wediscuss

thevariousoptionsandpresentanalyticalarguments

toeliminate

certaindesigncombinations.

5VS/VDSwitchDesignOptions

5.1

VCRate

MeasurementTechniques

Wehavepresentedfourwaysof

�ndingthetheVC'scurrentrate

insection4.1,twoof

them

used

declaredratesandtwoof

them

measuredtheactual

sourcerate.Weshow

that

measuringsource

ratesisbetterthan

usingdeclaredratesfortworeasons.

10

First,thedeclaredVCrateofaloop

naivelyistheminimumofbottleneckratesof

downstream

loops

only.It

does

not

consider

thebottleneckratesof

upstream

loops,andmay

ormay

not

consider

thebottleneckrate

ofthe�rstlinkofthenextloop.Measurementallowsbetterestimationofload

when

thetra�

cisnot

regular.

Second,theactualrate

oftheVCmay

belower

than

thedeclaredACRoftheVCdueto

dynam

ic

changesin

bottleneckratesupstream

ofthecurrentswitch.Thedi�erence

inACR

andVCrate

willremainatleastas

longas

thetimerequired

fornew

feedbackfrom

thebottleneckin

thepath

toreachthesourceplusthetimeforthenew

VCrate

tobeexperiencedat

theswitch.Thesum

of

thesetwodelay

componentsiscalled

the\feedbackdelay."

Dueto

feedbackdelay,itispossiblethat

thedeclaredrate

isastalevalueat

anypointof

time.

Thisisespeciallytruein

VS/V

Dswitches

whereper-VCqueues

may

controlsourceratesto

values

quitedi�erentfrom

theirdeclaredrates.

Further,themeasuredsourcerate

caneasily

becalculatedin

aVS/V

Dswitch

since

thenecessary

quantities

(number

ofcellsandtimeperiod)aremeasuredas

partof

oneof

thesourceendsystem

rules(SESRule5)

[1,2,10].

5.2

InputRate

measurementtechniques

Asdiscussed

earlier,theinputrate

canbemeasuredas

thesum

oftheinputratesof

VCsto

the

per-VCqueues

ortheaggregateinputrate

totheper-classqueue.

Thesetworatescanbedi�erent

because

theinputrate

totheper-VCqueues

isat

thepreviousloop'srate

whiletheinputto

the

per-classqueueisrelatedto

thenextloop'srate.Figure10

show

sasimplecasewheretwoadjacent

loopscanrunat

very

di�erentrates(10Mbpsand100M

bps)foronefeedbackdelay.

Figure

10:Twoadjacentloopsmay

operateat

very

di�erentratesforonefeedbackdelay

5.3

CombinationsofVCrate

andinputrate

measurementoptions

Table1summarizes

theoption

combinationsconsideringthefact

that

twoadjacentloopsmay

run

atdi�erentrates.

Thetableshow

sthat

fourof

thesecombinationsmay

worksatisfactorily.The

11

other combinations use inconsistent information and hence may either overallocate rates leading to

unconstrained queues or result in unnecessary oscillations. We can eliminate some more cases as

discussed below.

Table 1: Viable combinations of VC rate and input rate measurement

# VC Rate � VC rates (Mbps) Input Rate Input Rate Design

Method Method Value (YES/NO)

1. From FRM1 10 � per-VC 10 YES

2. From FRM1 10 per-class 10-100 NO

3. From FRM2 100 � per-VC 10 NO

4. From FRM2 100 per-class 100 YES

5. At per-VC queue 10 � per-VC 10 YES

6. At per-VC queue 10 per-class 10-100 NO

7. At per-class queue 100 � per-VC 10 NO

8. At per-class queue 100 per-class 100 YES

The above table does not make any assumptions about the queue lengths at any of the queues

(per-VC or per-class). For example, when the queue lengths are close to zero, the actual source rate

might be much lower than the declared rate in the FRMs leading to overallocation of rates. This

criterion can be used to reject more options.

The performance of one such rejected case is shown in Figure 11 (corresponding to row 4 in Table 1).

The con�guration used has two ABR in�nite sources and one high priority VBR source contending

for the bottleneck link's (LINK1) bandwidth. The VBR has an ON/OFF pattern, where it uses

80% of the link capacity when ON. The ON time and the OFF time are equal (20 ms each). The

VS/VD switch overallocates rates when the VBR source is OFF. This leads to ABR queue backlogs

when the VBR source comes ON in the next cycle. The queue backlogs are never cleared, and

hence the queues diverge. In this case, the fast response of VS/VD is harmful because the rates are

overallocated.

In this study, we have not evaluated row 5 of the table (measurement of VC rate at entry to the

per-VC queues). Hence, out of the total of 8 combinations, we consider two viable combinations:

row 1 and row 8 of the table. Note that since row 8 uses source rate measurement, we expect it

12

(a) ACR

(b) Queue Lengths

(c) Cells Received

(d) Con�guration

Figure 11: 2-source+VBR con�guration. Unconstrained queues due to overallocation.

to show better performance. This is substantiated by our simulation results presented later in the

paper.

5.4 E�ect of Link Congestion Control Actions

In a network with non-VS/VD switches only, the bottleneck rate needs to reach the sources before

any corresponding load change is seen in the network. However, a VS/VD switch can enforce the

new bottleneck rate immediately (by changing the ACR of the per-VC queue at the VS). This

rate enforcement a�ects the utilization of links in the next loop. Hence, the VS/VD link congestion

control actions can a�ect neighboring loops. We have enumerated three options in an earlier section.

We note that the second option (\next loop only") does not work because the congestion information

is not propagated to the sources of the congestion (as required by the standard [1]). This leaves

us with two alternatives. The third option (\both loops") is attractive because, when ACR2 is

updated, the switches in the next loop experience the load change faster. Switch algorithms may

save a few iterations and converge faster in these cases.

Figure 12 shows the fast convergence in a parking lot con�guration when such a VS/VD switch is

13

(a) ACR

(b) Queue Lengths

(c) Con�guration

Figure 12: Parking lot con�guration. Illustrates fast convergence of the best VS/VD option.

used. The parking lot con�guration (Figure 12(c))consists of three VCs contending for the Sw2-Sw3

link bandwidth. Link lengths are 1000 km and link bandwidths are 155.52 Mbps. The target rate

of the ERICA algorithm was 90% of link bandwidth i.e., 139.97 Mbps. The round trip time for the

S3-D3 VC is shorter than the round trip time for the other two VCs. The optimum allocation by

ERICA for each source is 1/3 of the target rate on the Sw2-Sw3 (about 46.7 Mbps). Figure 12(a)

shows that the optimum value is reached at 40 ms. Part (b) of the �gure shows that the transient

queues are small and that the allocation is fair.

�g:example-conv

5.5 Link Congestion and Allocated Rate Update Frequency: Viable

Options

The allocated rate update has three options:

a) update upon BRM receipt (in VS) and enter the value in a table to be used when an FRM is

turned around,

b) update upon FRM turnaround (at VD) and no action at VS,

c) update both at FRM (VD) and at BRM (VS) without use of a table.

14

Thelastoption

recomputestheallocatedratealargernumberoftimes,butcanpotentiallyallocate

ratesbetterbecause

wealwaysuse

thelatestinform

ation.

Theallocatedrate

updateandthee�ectsoflinkcongestionactionsinteract

asshow

nin

Figure13.

The�gure

show

satree

wherethe�rstlevelconsidersthelinkcongestion(2

options),i.e.,whether

thenextloop

isalso

a�ectedor

not.Thesecondlevelliststhethreeoptionsfortheallocatedrate

updatefrequency.Theviableoptionsarethosehighlightedin

boldat

theleaf

level.

Figure

13:Linkcongestionandallocatedrate

update:

viableoptions

Other

optionsarenot

viablebecause

ofthefollow

ingreasons.

Inparticular,ifthelinkcongestion

does

not

a�ectthenextloop,theallocatedrate

updateat

theFRM

turnaroundis

allthat

is

required.Theallocatedrate

attheBRM

isredundantin

thiscase.Further,ifthelinkcongestion

a�ects

thenextloop,then

theallocatedrate

updatehas

tobedoneon

receivingaBRM,so

that

ACRcanbechangedat

theVS.Thisgivesustwopossibilitiesas

show

nin

the�gure

(BRM

only,

andBRM+FRM).

Hence,wehavethreeviablecombinationsoflinkcongestionandtheallocatedrateupdatefrequency.

Asummaryof

allviableoptions(a

totalof

6)islisted

inTable2.

Table2:

Summaryof

viableVS/V

Ddesignalternatives

Option

#VCRate

InputRate

LinkCongestion

AllocatedRate

Method

Measurementpoint

E�ect

Updated

at

41From

FRM1

per-VC

previousloop

only

FRM1only

52measuredat

per-classQ

per-class

bothloops

FRM1only

329

From

FRM1

per-VC

bothloops

FRM1only

340

measuredat

per-classQ

per-class

bothloops

FRM1andBRM2

393

From

FRM1

per-VC

bothloops

BRM2only

404

measuredat

per-classQ

per-class

bothloops

BRM2only

Thenextsectionevaluatestheperform

ance

oftheviableVS/V

Ddesignoptionsthrough

simulation.

15

6Perform

anceEvaluationofVS/VDDesignOptions

6.1

Metrics

Weuse

fourmetrics

toevaluatetheperform

ance

ofthesealternatives:

�Response

Tim

e:isthetimetakento

reachnearoptimalbehavioron

startup.

�ConvergenceTim

e:isthetimeforrate

oscillationsto

decrease(timeto

reachthesteady

state).

�Throughput:

Totaldatatransferredper

unittime.

�Maxim

um

Queue:Themaximum

queuebeforeconvergence.

Thedi�erencebetweenresponsetimeandconvergence

timeisillustratedinFigure14.Thefollow

ing

sectionspresentsimulation

resultswithrespectto

theabovemetrics.Notethat

wehaveusedgreedy

(in�nite)

tra�

csources

inoursimulations.Wehavestudiedthealgorithmicenhancementsin

non-

VS/V

Dswitchesfornon-greedysourcesinreference

[8].Weexpectconsistentresultsforsuch

tra�

c

when

thebestimplementation

option

(see

below

)isused.

Figure

14:Response

timevsConvergence

time

6.1.1

Response

Tim

e

WithoutVS/V

Dallresponse

times

arecloseto

theround-tripdelay.WithVS/V

D,theresponse

times

arecloseto

thefeedbackdelay

from

thebottleneck.Since

VS/V

Dreducestheresponse

time

duringthe�rstroundtrip,itisgoodforlongdelay

paths.

Thequickresponse

time(10msin

the

parkinglotcon�guration

whichhas

a30

msroundtriptime)was

illustratedpreviouslyinFigure12.

Responsetimeisalso

importantforbursty

tra�

clike

TCP�letransferoverATM

which\startsup"

atthebeginningof

everyactive

period(when

theTCPwindow

increases)

afterthecorresponding

idleperiod[9,10].

16

6.1.2

Throughput

Thenumber

ofcellsreceived

atthedestinationisameasure

ofthethroughputachieved.These

values

arelisted

inTable

3.Thetoprow

isalist

ofthecon�guration

codes

(thesecodes

are

explained

inTable2.

The�nalcolumnliststhethroughputvaluesforthecase

when

anon-VS/V

D

switch

isused.The2source+

VBR

andtheparkinglotcon�gurationshavebeenintroducedin

earliersection.

Theupstream

bottleneckcon�guration

show

nin

Figure

15has

abottleneckat

Sw1where15

VCs

sharetheSw1-Sw2link.AsaresulttheS15-D15

VCisnot

capableofutilizingitsbandwidth

share

attheSw2-Sw3link.Thisexcessbandwidth

needsto

beshared

equallybytheothertwoVCs.The

tableentryshow

sthenumber

ofcellsreceived

atthedestinationforeither

theS16-D16

VCor

the

S17-D17

VC.

Inthe2source+

VBRandtheupstream

bottleneckcon�gurations,thesimulation

was

runfor400ms

(thedestinationreceivesdatafrom

time=15

msthrough

400ms).In

theparkinglotcon�guration,

thesimulation

was

runfor200m

s.

Figure

15:Upstream

bottleneckcon�guration

Table3:

Cellsreceived

atthedestinationper

sourcein

Kcells

VS/V

DOption

#!

AB

CD

EF

NoVS/V

D

Con�guration

#

2source+VBR

3131

32.5

3432

3330

Parkinglot

2222

2320.5

2320.5

19.5

Upstream

bottleneck

6161

6160

6161

62

17

As we compare the values in each row of the table, we �nd that, in general, there is little di�erence

between the alternatives in terms of throughput. However, there is a slight increase in throughput

when VS/VD is used over the case without VS/VD switch.

6.1.3 Convergence Time

The convergence time is a measure of how fast the scheme �nishes the transient phase and reaches

steady state. It is also sometimes called \transient response." The convergence times of the various

options are shown in Table 4. The \transient" con�guration mentioned in the table has two ABR

VCs sharing a bottleneck (like the 2 source + VBR con�guration, but without the VBR VC). One

of the VCs comes on in the middle of the simulation and remains active for a period of 60 ms before

going o�.

Table 4: Convergence time in ms

VS/VD Option # ! A B C D E F No VS/VD

Con�guration #

Transient 50 50 65 20 55 25 60

Parking lot 120 100 170 45 125 50 140

Upstream bottleneck 95 75 75 20 95 20 70

Observe that the convergence time of VS/VD option D (highlighted) is the best. Recall that this

con�guration corresponds to measuring the VC rate at the entry to the per-class queue, input rate

measured at the per-class queue, link congestion a�ecting both the next loop and the previous loop,

the allocated rate updated at both FRM1 and BRM2.

6.1.4 Maximum Transient Queue Length

The maximum transient queues gives a measure of how askew the allocations were when compared

to the optimal allocation and how soon this was corrected. The maximum transient queues are

tabulated for various con�gurations for each VS/VD option and for the case without VS/VD in

Table 5.

The table shows that VS/VD option D has very small transient queues in all the con�gurations

18

Table 5: Maximum queue length in Kcells

VS/VD Option # ! A B C D E F No VS/VD

Con�guration #

2 Source + VBR 1.2 1.4 2.7 1.8 2.7 1.8 2.7

Transient 1.4 1.1 1.4 0.025 1.3 1.0 6.0

Parking lot 1.9 1.9 1.4 0.3 3.7 0.35 2.0

Upstream bottleneck 0.025 0.08 0.3 0.005 1.3 0.005 0.19

and the minimum queues in a majority of cases. This result, combined with the fastest response

and near-maximum throughput behavior con�rms the choice of option D as the best VS/VD im-

plementation.

Observe that the queues for the VS/VD implementations are in general lesser than or equal to

the queues for the case without VS/VD. However, the queues reduce much more if the correct

implementation (like option D) is chosen.

7 Conclusions

In summary:

� VS/VD is an option that can be added to switches which implement per-VC queueing. The

addition can potentially yield improved performance in terms of response time, convergence

time, and smaller queues. This is especially useful for switches at the edge of satellite networks

or switches that are attached to links with large delay-bandwidth product. The fast response

and convergence times also help support bursty tra�c like data more e�ciently.

� The e�ect of VS/VD depends upon the switch algorithm used and how it is implemented

in the VS/VD switch. The convergence time and transient queues can be very di�erent for

di�erent VS/VD implementations of the same basic switch algorithm. In such cases the fast

response of VS/VD is harmful.

� With VS/VD, ACR and actual rates are very di�erent. The switch cannot rely on the RM

cell CCR �eld. We recommend that the VS/VD switch and in general, switches implementing

per-VC queueing measure the VC's current rate.

19

� The sum of the input rates to per-VC VS queues is not the same as the input rate to the link.

It is best to measure the VC's rate at the output of the VS and the input rate at the entry to

the per-class queue.

� On detecting link congestion, the congestion information should be forwarded to the previ-

ous loop as well as the next loop. This method reduces the convergence time by reducing

the number of iterations required in the switch algorithms on the current and downstream

switches.

� It is best for the the rate allocated to a VC to be calculated both when turning around FRMs

at the VD as well as after receiving BRMs at the next VS.

8 Future Work

The VS/VD provision in the ABR tra�c management framework can potentially improve perfor-

mance of bursty tra�c and reduce the bu�er requirements in switches. The VS/VD mechanism

achieves this by breaking up a large ABR loop into smaller ABR loops which are separately con-

trolled. However, further study is required in the following areas:

� E�ect of VS/VD on bu�er requirements in the switch.

� Scheduling issues with VS/VD.

� E�ect of di�erent switch algorithms in di�erent control loops, and di�erent control loop

lengths.

� E�ect of non-ABR clouds and standardization issues involved.

� E�ect of using switch algorithms speci�cally designed to exploit the per-VC queueing policy

required in VS/VD implementations.

References

[1] ATM Forum, \ATM Tra�c Management Speci�cation Version 4.0," April 1996, available as

ftp://ftp.atmforum.com/pub/approved-specs/af-tm-0056.000.ps

20

[2] Raj Jain, Shiv Kalyanaraman, Rohit Goyal, Sonia Fahmy, \Source Behavior for ATM ABR

Tra�c Management: An Explanation," IEEE Communications Magazine, November 19962.

[3] Kerry Fendick, \Evolution of Controls for the Available Bit Rate Service," IEEE Communica-

tions Magazine, November 1996.

[4] L. Roberts, \Enhanced PRCA (Proportional Rate-Control Algorithm)," AF-TM 94-0735R1,

August 1994.

[5] K. Siu and T. Tzeng, \Intelligent congestion control for ABR service in ATM networks,"

Computer Communication Review, Volume 24, No. 5, pp. 81-106, October 1995.

[6] L. Kalampoukas, A. Varma, K. K. Ramakrishnan, \An e�cient rate allocation algorithm for

ATM networks providing max-min fairness," Proceedings of the 6th IFIP International Con-

ference on High Performance Networking, September 1995.

[7] Y. Afek, Y. Mansour, and Z. Ostfeld, \Phantom: A simple and e�ective ow control scheme,"

Proceedings of the ACM SIGCOMM, August 1996.

[8] R. Jain, S. Kalyanaraman, R. Goyal, S. Fahmy, and R. Viswanathan, \The ERICA Switch

Algorithm for ABR Tra�c Management in ATM Networks, Part I: Description" IEEE Trans-

actions on Networking, submitted.

[9] Anna Charny, Gunter Leeb, Michael Clarke, \Some Observations on Source Behavior 5 of the

Tra�c Management Speci�cation," AF-TM 95-0976R1, August 1995.

[10] Shivkumar Kalyanaraman, Ray Jain, Sonia Fahmy, Rohit Goyal, \Use-It-or-Lose-It (UILI)

Policies for the ABR Service in ATM Networks," Computer Networks and ISDN Systems,

submitted.

2All our papers and ATM Forum contributions are available through http://www.cis.ohio-state.edu/~jain

21