0 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.?, NO.?, MONTH? Adaptve Call Admsson and Bandwdth Control n DVB-RCS Systems Maro Marchese and Maurzo Mongell Abstract: The paper presents a control archtecture amed at mplementng bandwdth optmzaton combned wth Call Admsson Control (CAC) over a DVB Return Channel Satellte Termnal (RCST) under Qualty of Servce (QoS) constrants. The approach can be appled n all cases where traffc flows, comng from a terrestral porton of the network, are merged together wthn a sngle DVB flow, whch s then forwarded over the satellte channel. The paper ntroduces the archtecture of data and control plane of the RCST at layer 2. The data plane s composed of a set of traffc buffers served wth a gven bandwdth. The control plane proposed n ths paper ncludes a Layer 2 Resource Manager (RM), whch s structured nto Decson Makers (DM), one for each traffc buffer of the data plane. Each DM contans a vrtual queue, whch exactly duplcates the correspondng traffc buffer and performs the actons to compute the mnmum bandwdth need to assure the QoS constrants. After computng the mnmum bandwdth through a gven algorthm (n ths vew the paper reports some schemes taken n the lterature whch may be appled), each DM communcates ths bandwdth value to the RM, whch allocates bandwdth to traffc buffers at the data plane. Real bandwdth allocatons are drven by the nformaton provded by the DMs. Bandwdth control s lnked to a CAC scheme, whch uses current bandwdth allocatons and peak bandwdth of the call enterng the network to decde admsson. The performance evaluaton s dedcated to show the effcency of the proposed combned bandwdth allocaton and CAC. Index Terms: call admsson, DVB-RCS/S2, measurement-based equvalent bandwdth, QoS mappng. A. Technologcal scenaro I. INTRODUCTION Ths paper deals wth a satellte network, based on the DVB- RCS standard [1, 2], composed of a GEO statonary bent-ppe satellte, Return Channel Satellte Termnals (RCSTs), and a Network Control Center (NCC) that s connected to the Internet (Fg. 1). Local Area Networks (LANs) may be connected to RCSTs. RCSTs are fxed and use the Return Channel va Satellte (RCS). The NCC provdes control and montorng functons and manages network resources allocaton to RCSTs. DVB- S s used for the forward lnk (from NCC to RCSTs, wth a data rate hgher than 10 Mbps) and DVB-RCS s employed for the return lnk (from RCSTs to NCC, wth a data rate around 2 Mbps). In DVB-S, the Mult Protocol Encapsulaton (MPE) provdes segmentaton and reassembly of IP packets, thus provdng the MPEG2 stream whose packets sze amounts of 188 Manuscrpt receved x x, 2010; revsed x x, 2010, approved for publcaton by x x, Dvson x x Edtor, x x, x. DIST - Department of Communcatons, Computer and Systems Scence Unversty of Genova, Va Opera Pa 13, 16145, Genova, Italy, {maro.marchese, maurzo.mongell @unge.t} 1229-2370/03/$10.00 c 2003 KICS Fg. 1. DVB-RCS-S2 communcaton system. bytes. The transmsson s performed through channel codng and modulaton. The DVB-RCS ar nterface s based on Mult Frequency - Tme Dvson Multple Access. The NCC assgns a group of slots, characterzed by frequency, bandwdth, start tme and duraton, to each RCST, and communcates resource allocatons to RCSTs through the Termnal Burst Tme Plan. DVB-RCS makes use of 5 DVB classes, each mplemented through a dedcated queue at DVB layer, and 5 correspondng resource allocaton types: () Contnuous Rate Assgnment (CRA); () Rate-Based Dynamc Capacty (RBDC); () Volume-Based Dynamc Capacty (VBDC), where an RCST dynamcally and cumulatvely requests the total number of slots needed to dle ts queue; (v) Absolute Volume-Based Dynamc Capacty (AVBDC), where an RCST dynamcally requests the number of slots, but requests are not cumulatve; (v) Free Capacty Assgnment (FCA). VBDC and AVBDC, whch mply tmevaryng bandwdth requests, and CRA, whose requrements affect the overall avalable bandwdth that can be dynamcally allocated, are meanngful for ths paper. B. State of the Art and Motvatons The problem of bandwdth allocaton among the RCSTs to satsfy QoS levels naturally arses n DVB envronments, especally for the return channel where bandwdth s a scarce resource. Ths, n turn, leads to the problem of QoS mappng of IP flows over the DVB classes (see [3], [4] and references theren). Ths topc receved the attenton of the satellte communty n the last years for what concerns research projects (e.g., [5], [6]) and scentfc lterature (an excellent overvew can be found n [7]). The nature of the problem recalls the prncples of "adaptve feedback control", as DVB classes dynamcally ask and re-
MARCHESE et al.: TITOLO... 1 lease bandwdth resources on the bass of measures of the traffc flows. The overall am of dynamc bandwdth control s to avod wastng resources. Overprovsonng of a-pror allocatons s hghly neffcent n tme-varyng condtons. Several works address the dynamc bandwdth control problem n DVB networks (see [8] and references theren), by explotng optmal control methodologes, e.g., [9], and TCP adaptatons to the satellte, e.g., [10]. A pecularty of the DVB technology, not addressed by the mentoned lterature, s that bandwdth allocaton must be mplemented when IP flows wth dfferent statstcs and QoS constrants are merged nto a smaller set of DVB streams at layer 2 [4, 7]. Ths acton s called vertcal QoS mappng. Fg. 2. RCST Layer 2. C. Am of ths paper The underlyng dea of ths paper s to desgn control blocks and actons of a combned Call Admsson Control (CAC) Bandwdth Allocaton scheme, mplemented wthn the RCST Layer 2 Resource Manager (RM). The desgn s addressed by nvestgatng theoretcal aspects and protocol rules, and t s heavly based on the concept of vrtual queues. Gven a real traffc buffer wthn each RCST, a vrtual queue s another queue, mrrorng the real traffc one, whch receves data and can perform measures and actons wthout nterferng wth routne traffc forwardng. The words "buffer" and "queue" are used ndfferently n ths paper. The specfc algorthm to compute the mnmum bandwdth, commonly called "equvalent bandwdth" n the scentfc lterature, s performed through schemes taken n the lterature and based on traffc measures. The dea of drvng the CAC wth measurement-based equvalent bandwdth s derved from [11, 12]; the computaton s extended to heterogeneous condtons n [13] and to the presence of nosy wreless channels [4]. A huge quantty of lterature addresses the concept of bandwdth and CAC control, see, e.g., [14, 15]. Addtonally, the soluton proposed n ths paper s amed at practcal mplementaton and use n the feld. For ths motvaton not only scentfc lterature s referenced but also actve patents such as [16], whch uses the concept of vrtual queue to drve bandwdth allocaton as n ths paper, and [17], whch may be applcable as a possble control law for the computaton of the bandwdth need. Bandwdth avalablty over satelltes s often tme varant both because of fadng and of termnal moblty. Ths aspect s mathematcally modelled n ths paper through a multplcatve factor that decreases the bandwdth avalable for data transmsson. The paper s structured as follows: the next secton contans the proposed archtecture of the RM, whch ncludes vrtual queues and Decson Makers (DMs). Secton III ntroduces the acton of vrtual queue lnked to bandwdth allocaton and descrbes the role of DMs. Secton IV presents the bandwdth allocaton acton at RM and the CAC rule, and dscusses the possble mpact of tme varant bandwdth avalablty over bandwdth allocaton. Secton V reports some algorthms taken n the lterature and used wthn DMs. Secton VI extends the proposed archtecture and control soluton when more than one flow s multplexed over one DVB queue. Secton VII shows the performance evaluaton and secton VIII the conclusons. II. ARCHITECTURE OF RCST LAYER 2 RESOURCE MANAGER (RM) Fg. 2 shows data and control planes of the RCST. Gven N traffc flows at layer 2, the data plane s modeled as a set of traffc buffers, one for each flow, (f1,..., fn ), served wth rate θ 1,..., θ N, respectvely. The layer 2 receves the traffc flows from the layer 3, whch s not the object of hs paper. A proposal concernng formal communcaton between layer 3 and layer 2 s reported n [18]: the IP protocol stack s dvded nto lower layers (layer 2 and 1), called Satellte Dependent (SD) layers, and upper layers (layer 3-IP and above layers), called Satellte Independent (SI) layers. The nterface between SI and SD layers (n practce the nterface between IP and layer 2) s defned through an nterface called SI-SAP (Satellte Independent Servce Access Pont), whch provdes a set of communcaton prmtves for IP - layer 2 communcaton and assures the separaton of SI, ndependent of satellte technology, and SD layers, strctly dependent on the used satellte technology. In ths formal context, IP traffc flows, as well as IP flows performance requrements, f any, can flow through the SI-SAP and must be mapped at layer 2 (DVB, n ths paper). The control plane contans the RM, whch s composed of N DMs, whose role, specfed n detal n the followng, s to compute an estmaton (θ1 mn,..., θn mn ) of the mnmum bandwdth necessary at traffc buffers to provde a gven qualty of servce (QoS 1,..., QoS N ). After computng the mnmum bandwdths, the vector (θ1 mn,..., θn mn ) s forwarded to the Bandwdth Allocator block that decdes and communcates to traffc buffers real bandwdth allocatons (θ 1,..., θ N ). Real bandwdth allocatons should depend on the values of the mnmum bandwdth. Mathematcally: θ 1 = f(θ1 mn ),..., θ N = f(θn mn ). A proposal concernng functon f( ) s reported n the followng. III. DECISION MAKERS: VIRTUAL QUEUES AND MINIMUM BANDWIDTH COMPUTATION The archtecture of generc -th DM s shown n Fg. 3. Each traffc flow (f1,..., fn ) s dvded nto two parts. One s drected towards the correspondng traffc buffer to be forwarded to the physcal layer as shown n Fg. 2. The other part s sent to another buffer, the vrtual queue, whch s an exact mrror of the traffc one but t does not nterfere wth routnary forwardng operatons. Vrtual queues are one of the most mportant con-
2 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.?, NO.?, MONTH? Fg. 3. DM archtecture. cepts wthn the block schemes ntroduced n ths paper. Traffc buffers and vrtual queues work n parallel but vrtual queues are served wth the estmated mnmum rates (θ1 mn,..., θn mn ), computed by the Estmated Mnmum Bandwdth Computaton block n each DM. DM computes θ mn by followng a set of steps. A sequence k = 1, 2,... of observaton horzons for DM s defned (OH (k)), durng whch the vrtual queue s montored n approprate nstants of tme. These nstants and montorng nformaton compose an nformaton vector I (k) for each OH (k). I (k) drves the servce rate computaton of queue at tme k + 1, together wth the prevous allocatons up untl a tme depth d, thus generatng θ mn (k + 1). θ mn (k+1) = F ( θ mn (k),..., θ mn (k d), I (k),..., I (k d) ) (1) Fg. 4. Estmated mnmum bandwdth computaton by DM. horzons. More refned stablzaton condtons may be used and reasonably appled n the context of the paper. The replcaton process of the packets towards vrtual queue (together wth the computatons to obtan θ mn ) may requre the applcaton of a dedcated chp n case of computatonal lmtaton of the natve hardware structure of the RCST. After evaluatng the stablzaton of θ mn, each DM transmts the mnmum bandwdth values to RM Bandwdth Allocator (see Fg. 2), whch acts as ndcated n the next secton. (k), and the QoS threshold values QoS : e (, k) = (QoS QoS OH (k)) 2. A possble smplfcaton of (1) s contaned n (2), where θ mn (k + 1) s computed by usng only nformaton at nstant k. The actons are repeated by DM durng each OH (k), k = 1, 2,... An example of nformaton vector may be represented by the vector of dfferences e 1,..., e N between the QoS levels measured over the vrtual queues durng the observaton horzons OH (k) and denoted by QoS OH θ mn (k + 1) = F (θ mn (k), e (, k)) (2) Fg. 4 shows the steps to get θ mn (k + 1) takng (2) as a reference. Usually, the temporal dmenson of the OH ( ) s n the range [30, 360] s, dependng on the specfc applcatons to be montored wthn queue. Examples of algorthms to compute θ mn are reported n secton V. Actually, the specfc mnmum bandwdth computaton s not part of the novelty of ths paper but t may be nterestng to have an dea of possble solutons leavng full detals to specfc papers n the lterature. After gettng θ mn each DM mplements a scheme to evaluate the stablzaton of θ mn computaton. The steady state of θ mn, [1,..., N ] s captured under the followng condton: θ mn (k + 1) θ mn (k) ɛ, where ɛ s the stablzaton threshold for queue. A reasonable practcal value of ɛ mght be: ɛ = 1 10 θmn (k). Ths means θ mn s n steady state f t shows small oscllatons between two consecutve observaton IV. BANDWIDTH ALLOCATION AT RM AND CAC A. RM Bandwdth Allocator The real bandwdth θ, allocated to queue, s changed over tme by followng ndcatons comng from DM and dependng on the value of θ mn. Ths paper proposes a possble acton but other solutons may be appled. RM, on the bass of the value of θ mn, assumes three possble states for queue : a) no acton requred ; b) mmnent congeston ; 3) bandwdth release. Ideally the 3 states can be defned by usng only the value of θ mn : f θ < θ mn f θ = θ mn f θ > θ mn not enough allocated bandwdth - mmnent congeston; mnmum allocated bandwdth - no acton requred; too much allocated bandwdth - bandwdth release. Operatvely, the proposal s too hard and can generated unexpected bandwdth oscllatons wth consequent mpact on real traffc. A smoother soluton s preferred n practce: the mentoned states are defned by 2 thresholds (1 up ) and (1 + down ) as follows: f θ < (1 up ) θ mn not enough allocated bandwdth - mmnent congeston; f (1 up ) θ mn θ (1 + down ) θ mn mnmum allocated
MARCHESE et al.: TITOLO... 3 Fg. 5. Queue possble states. bandwdth - no acton requred; f θ > (1 + down ) θ mn too much allocated bandwdth - bandwdth release. Fg. 5 graphcally shows the concepts ntroduced above: f the allocated bandwdth θ s close enough to θ mn, no acton s requred; f θ value s too much below θ mn, bandwdth s underprovsoned and congeston may happen n the next future; f θ value s too much above θ mn, bandwdth s overprovsoned and may be released. Reasonable practcal values of up and down, not necessarly equal, may be n the range [0.05,0.3]. B. Call Admsson Control (CAC) The bandwdth update provded by the RM Bandwdth Allocator descrbed above s lmted by the maxmum avalable bandwdth C at a specfc RCST. Ths value s provded by the NCC, shown n Fg. 1. In ths case, referrng to the specfc RCST n Fg. 2, whch mplements N queues, there s the followng constrant: N θ C. The maxmum bandwdth value =1 allocated by the NCC may be seen also dvded for sngle queue wthn the RCST, f needed for control reasons. For example, referrng agan to the specfc RCST n Fg. 2, a maxmum bandwdth C mght be forecast for queue. The constrant s N C =1 = C. Obvously, n ths case, θ C, = 1,..., N. CAC s mplemented wthn RM and must consder both the currently used bandwdth θ and the maxmum avalable bandwdth, ether C or C, as explaned above. The scheme used n ths paper s very smple but effectve f appled together wth the bandwdth allocaton scheme proposed n the prevous subsecton. Wthout makng any per queue maxmum bandwdth allocaton, so referrng to constrant N θ C, an ncomng connecton wth =1 peak bandwdth p s accepted at queue f N θ + p C. =1 Alternatvely, assgnng a maxmum bandwdth C at each queue, = 1,..., N, so referrng to the constrant set θ C, = 1,..., N, a new connecton, wth peak bandwdth p, s accepted at queue f θ + p C. Even f CAC s performed on the bass of the peak bandwdth, overprovsonng s only temporary because θ s updated by RM as shown before. C. Tme varant bandwdth avalablty: tacklng fadng and termnal moblty Satellte bandwdth avalablty may be tme varant for dfferent motvatons. One of them s represented by fadng. Obvously bandwdth varablty s not the only effect of fadng but ths paper lmts ts scope to t. The concept may be also seen from another vewpont: ths paper models fadng as a bandwdth reducton. In practce, dfferent fadng classes are defned, correspondng to combnatons of channel bt and codng rate that gve rse to redundancy factors ξ level (t), level = 1, 2,... (ξ level (t) 1.0). ξ level (t) represents the rato between the Informaton Bt Rate (IBR) n clear sky and the IBR n specfc workng condtons. The correspondng bandwdth reducton factor s defned as: φ = 1 ξ level (t). The bandwdth reducton at the queue, denoted by θ real (t), can be computed as θ real (t) = φ θ (t),. Concernng termnal moblty, t s a recent research topc, of man nterest for ndustry. One of the man problems s handover, whch occurs when a moble RCST changes several satellte beams over tme. Research projects have been recently launched concernng moblty over satelltes [19, 20]. Durng the movements, physcal effects, such as doppler and multpath fadng, may degrade the performance at the physcal level [21]. Crtcal communcaton perods mply ether sgnal blockages, whch generate complete outage and no bandwdth avalablty, or due to shadowng. The countermeasures, for both the forward and return lnks, are: lnk layer- FEC (such as raptor codes), manly to tackle shadowng, and Proactve Retransmsson (PR), to tackle total outage. The former may be modeled through generc multplcatve bandwdth reducton factors, smlarly as done for fadng. The latter conssts of freezng the transmtter, thus bufferng the ncomng data durng outage perods. The approach proposed n ths paper can be nvolved n ths mechansm as follows. When an outage s foreseen ( smart mode proposed by [22] suggests that a moble RCST can estmate ts poston over tme by usng GPS and so can try foreseeng outage perods) over a gven perod of tme, the mnmum bandwdth estmaton algorthm, together wth data transmsson, should be freezed n order to avod useless bandwdth computatons durng data bufferng. After the outage perod, the estmaton of the mnmum bandwdth should be rentalzed wth the bandwdth value before the outage. V. ALGORITHMS TO COMPUTE THE MINIMUM BANDWIDTH θ mn Dfferent forms of the control law F ( ), appearng n formula (1) and (2), may be reasonably appled n the context of the paper to get θ mn. A. Reference Chaser Bandwdth Controller (RCBC) If the QoS of nterest s the Packet Loss Probablty (PLP) or the delay one could use Infntesmal Perturbaton Analyss to derve a gradent-based formulaton of the control law F ( ) as follows [4]: θ mn (k + 1) = θ mn e (, k) (k) + η k θ θ =θ mn (k) where η k s the gradent step sze; more specfcally, for the (3)
4 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.?, NO.?, MONTH? PLP case: e (, k) θ where θ =θ mn (k) = 2 ˆl (θ ) θ ˆl (θ ) θ θ = =θ mn (k) 1 T k N Tk bp=1 [ at bp T k (θ mn θ=θ mn (k) (k)) ll bp [ˆl (, k) ˆl ] (, k) T k (θ mn ] (k)) where ˆl (, k) s the measured loss rate of queue over the OH (k), ˆl (, k) s the target loss rate comng from the requred PLP value for queue (P LP ): ˆl (, k) = OH (k) P LP a (t)dt, a (t) s the measured nput rate of traffc class over the OH (k)), T k s the sze of OH (k). A busy perod (bp), n (5), s a perod of tme n whch the buffer s not empty. The quantty n brackets n (5) s the dfference between the last loss durng the busy perod bp and the startng tme of bp. A comparable gradent-based formulaton can be obtaned for the control of the delay performance (detals can be found n [23], and are not reported here for the sake of synthess). Ths technque s ndcated here as Reference Chaser Bandwdth Controller (RCBC) for loss control. B. PID control Other, more tradtonal, approaches for the control law F ( ) are possble. For example, Proportonal Integratve Dervatve (PID) control laws may be appled: θ mn (k + 1) = θ mn (k) + w k+1 (k + 1) e (, k + 1) + w k (k) e (, k) + w k 1 (k 1) e (, k 1) (6) where the weghts are the tunng parameters used to optmze the PID temporal behavor n dependence of the specfc applcaton of nterest. A huge amount of scentfc lterature exsts on PID applcatons and parameters optmzaton, also for the bandwdth allocaton case. The majorty of ndustral processes nowadays are stll regulated by PID controllers. Ths reveals the rch potental of ths smple control strategy for meetng varous specfcatons for a vast varety of practcal applcatons. The PID choce s mandatory for complcated metrcs, such as jtter, for whch no gradent formulatons are avalable. For both RCBC and PID cases, t s well known that f the nput rate processes of the buffer are ergodc and the qualty constrants do not vary over tme at least wthn convergence tmes, and other smple condtons are met (e,g the decreasng behavor of the gradent steps sze η k for the RCBC case), the above control laws converge to the exact value of θ mn ; t means that the requred QoS levels are satsfed wth the mnmum amount of bandwdth allocaton (.e., θ mn ). C. Equvalent bandwdth Other control laws, not drectly dependent on error e (, k), are applcable. An example s: (4) (5) θ mn (k + 1) = m (k) + 2ln(P LP ) ln(2π) σ (k) (7) The equaton above reports an Equvalent Bandwdth (EqB) technque applcable for the PLP case [11, 12]; m (k) and σ (k) are the average and standard devaton, respectvely, of the nput rate process of queue over OH (k) and P LP s the PLP requrement at queue. The jont control of PLP and delay wth EqB requre some heurstc adaptatons, such as the lmtaton of the maxmum delay by properly settng a-pror the queue sze to small values, see, e.g., [24]. D. Other approaches Other approaches are possble for the choce of the control law F ( ) and to get θ mn ; for example, some neural or fuzzy technques capable to support self-learnng adaptaton of the θ mn estmaton. Anyway, every measurement-based algorthm dedcated to support precse estmaton of θ mn can be appled wthn the framework reported n Fg. 4. Specfc attenton should be also devoted to the computatonal effort requred for the computaton. The effort of control laws mentoned above s low, partcularly for the RCBC and PID approaches. Concernng EqB, the computatonal burden depends on the specfc algorthm chosen to estmate mean and standard devaton parameters. VI. EXTENSION TO THE MULTICLASS CASE The secton s amed at extendng the control scheme presented n sectons from II to V to the case where more than one flow (comng from layer 3) s conveyed wthn one layer 2 flow (f, takng the flow as a reference) and, consequently, one layer 2 queue (, followng the same -th layer 2 reference). In practce, f s composed by more than one component flow, called traffc class n the followng where necessary to avod confuson. An example may be two IP flows, e.g. a VoIP and a vdeo flow, whch are multpled over a sngle DVB flow enterng a sngle DVB buffer. The DVB technology allows mappng dfferent IP traffc flows over dfferent DVB flows, thus mantanng the dstncton among traffc flows even f conveyed to a sngle DVB buffer. The tools are ether the Packet Identfer feld (13 bt) or the ATM vrtual channels dentfers, carred n the header of each DVB transport stream, n the forward or return lnk, respectvely. Agan, usng the example above: VoIP and vdeo flows are conveyed towards the same DVB buffer but they can stll be dfferentated. Ths dstncton may be reflected mathematcally by usng an addtonal ndex j, whch dentfes a sngle traffc class, wthn a layer 2 flow. Fg. 6 contans the representaton of the RCST Layer 2 of Fg. 2 n the multclass case. For each layer 2 flow (and correspondng buffer), the group of M traffc classes f,1 s multplexed over the flow f,..., f,j,..., f,m 1. Some quanttes defned up to now must be redefned by consderng also the ndex j. The qualty of servce threshold of each buffer are redefned per traffc class: (QoS,1,..., QoS,j,..., QoS,M ), as well as the qualty of servce measures n the observaton horzon OH (k), k = 1, 2,...: (QoS,1 OH,..., QoSOH,j,..., QoSOH,M ) and the errors (e,1 (, k),..., e,j (, k),..., e,m (, k)). In consequence, equatons (1) and ts smplfcaton (2), related to generc buffer may be rewrtten as:
MARCHESE et al.: TITOLO... 5 the multclass case, the equalty s no longer true. The equaton must be rewrtten as: ˆl,j (θ ) θ,j = θ,j =θ,j mn (k) 1 T k N Tk bp=1 [ at bp T k (θ,j mn (k)) ll bp T k (θ,j mn ] (k)) and represents an approxmaton ntroduced n [13]. (11) Fg. 6. RCST Layer 2 multclass case. Fg. 7. Estmated mnmum bandwdth computaton by DM multclass case. θ mn,j (k + 1) = F ( θ,j mn (k),..., θ,j mn (k d), I,j (k),..., I,j (k d) ) (8) θ,j mn (k + 1) = F (θ,j mn (k), e,j (, k)) (9) The estmated mnmum bandwdth θ mn (k + 1) related to buffer should be allocated as: θ mn (k + 1) = max[θ mn j,1 (k + 1),..., θ,j mn (k + 1),..., θ mn,m (k + 1)] (10) Consequently, the DM actons n Fg. 4 should be modfed as shown n Fg. 7. The CAC s not modfed by the ntroducton of traffc classes except for the defnton of the peak rate, whch may be defned per class. The algorthms to compute the estmated mnmum bandwdth shown n the prevous secton are stll applcable. From the formal vewpont t s suffcent addng the ndex j to the equatons n secton V. The only substantal modfcaton should be done to equaton (5) where, n VII. PERFORMANCE EVALUATION AND DISCUSSION The ams of the performance evaluaton are twofold. Frst, a performance analyss s made for the proposed bandwdth computaton algorthms; n order to stress the workng condtons, no CAC s appled and θ = θ mn (.e., the real queue s deally an exact replcaton of the vrtual one). Second, the proposed CAC s valdated under the proposed bandwdth allocaton scheme; a real fadng trace s also consdered n ths case. A. Bandwdth control wthout CAC The bandwdth control algorthms under nvestgaton are: RCBC, PID and EqB. An deal allocaton technque (Ideal) s also consdered for PLP control, whch explots a perfect knowledge of future packet arrvals (though t s not realstc, t can be done easly va smulaton). Ths knowledge offers support to a computaton of the exact bandwdth to assure a gven qualty of servce threshold. For the sake of synthess, PLP metrc s consdered here, thus dsregardng the delay metrc. There s just one traffc: VoIP. Accordng to ITU-T P.59, each source s an on-off process wth exponentally dstrbuted on and off tme duratons (mean 1.008 s and 1.587 s, respectvely) and peak bandwdth of 16 kbps. VoIP traffc enters an IP buffer whose length and servce rate (set by the traffc peak bandwdth) guarantee no packet loss rate. IP traffc s encapsulated nto DVB frames, thus generatng the process f, representng VoIP, shown n Fgs. 2 and 3; f enters the DVB buffer (62 DVB cells), where the VoIP loss rate s measured every OH. P LP = P LPV oip s set to 10 2 ; OH to 1 mnute. The number of VoIP sources s ncreased of 10 from 70 to 110 each 2124 s. The PID weghts are set here to 3.00, 1.50 and 1.25, respectvely. These values guarantee the best PID performance n the followng scenaro and were found through accurate smulaton nspecton va brute force analyss. RCBC gradent stepsze s set to 1.0 (no optmzaton of the gradent stepsze s provded). RCBC gradent descent s ntalzed by the VoIP average bandwdth of 70 sources and by consderng the ntroduced DVB overhead. Fg. 8 shows the resultng PLP at the end of each OH for all the technques and Fg. 9 the correspondng bandwdth allocatons. It s evdent from Fg. 8 that all the technques sometmes produce PLPs hgher than the threshold, except for EqB whch s always much below the threshold. Ideal assures the best performance. It must be noted that Ideal was optmzed by accurate smulaton nspecton: t was necessary to reduce ts OH to 57s (n place of the orgnal 1 mnute) because the Ideal computaton wth respect to future arrvals (regstered over OHs of 1 mnute each)
6 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.?, NO.?, MONTH? Fg. 8. VoIP scenaro: PLP. Fg. 9. VoIP scenaro: allocatons. Fg. 10. VoIP scenaro: Average performance. underestmates the necessary bandwdth to meet the target. The ratonale behnd ths behavor reles on the ntrnsc burstness of the sources, whch no approxmaton of the deal bandwdth (averaged over fnte tme perods) can match perfectly. The accuracy of the RCBC computaton s clear from Fg. 9. After two reallocaton steps, just at the begnnng of the smulaton, RCBC rate s smoothly changed over tme wth much hgher precson n comparson wth the consderable oscllatons provded by Ideal and PID. The smple observaton of Fgs. 8 and 9 suggests that RCBC reacts quckly to traffc changes also mnmzng bandwdth oscllatons. Ths has an mpact on the overall performance over the entre smulaton horzon. Quanttatve metrcs may help the nterpretaton of ths qualtatve behavor. Fg. 10 represents the average and standard devaton of PLP and bandwdth over the smulaton perod. Fg. 10 also shows the percentage of the OH perods where PLP s over threshold ( OverTrh ) and the average dfference between measured PLP and target ( AverageDffOTrh ), for all the technques consdered. Though PID exactly matches the target on average, t produces over-threshold PLPs for a relevant porton of tme. It means that ts bandwdth allocaton s not suffcent to assure the target performance over tme. EqB, on the other hand, overestmates the bandwdth need. The Ideal algorthm s almost perfect (only 12% of over-threshold PLPs). RCBC mnmzes the bandwdth effort whle assurng the average PLP closest to the target and mnmzng bandwdth oscllatons. Accordng to these results, a practcal applcaton of the decson schemes presented n secton III and IV can be ruled as follows. RCBC s used for computng θ mn (the RCBC convergence needs less than 5 OHs, see Fg. 9). At begnnng, θ s overprovsoned and later set to θ mn after 5 OHs. up and down n subsecton IV.A are both set to 0.1 because the RCBC allocaton varance s very low (0.007, see Fg. 10). B. Integraton wth CAC The applcaton of CAC s now consdered together wth bandwdth allocaton. A channel fadng condton s also consdered. There s just one buffer at layer 2. The maxmum avalable rate of the queue, ntroduced n the CAC subsecton of secton IV s C=2.0 Mbps; ts buffer sze amounts of 150 DVB cells. All the results gven n ths secton are based on averages of 10 ndependent smulatons, each of them of 8500 s. The fadng trace s appled after the frst 500 s perod. Together wth the VoIP traffc used n the prevous sub-secton, vdeo sources are also used. Both traffc classes are conveyed towards the sngle layer 2 buffer, so the multclass case ntroduced n secton VI must be consdered. Takng [24] as a reference, each vdeo source s modelled as an exponentally modulated on-off process, wth mean rate 0.25 Mbps and peak rate 0.7 Mbps and an average burst of 10.0 s. The actve tme of VoIP and vdeo sources are log-normally dstrbuted wth an average of 200 s. The actvaton processes are Posson arrval processes. Followng [11], varyng load condtons rangng from 0.5 to 5 are consdered, where the value of 1.0 corresponds to a mean source actvaton rate equal to 240 sources/hour (4 calls per mnute); an ncomng call s randomly set to be VoIP or vdeo wth dentcal probabltes. The performance metrcs under nvestgaton are the acheved PLP for each traffc class and lnk utlzaton under a gven load. Agan, the delay metrc s not consdered for the sake of smplcty; the maxmum achevable delay wth a full buffer (of 150 DVB cells wth rate of 2.0 Mbps) s 112 ms. It s a rare event wth a PLP lower than 10 2. The target PLPs are 10 2 for VoIP and 10 2 for vdeo. The employed fadng process, modeled through the φ factor ntroduced n secton IV, has been taken from [25], where real attenuaton samples are extracted from an expermental data set carred out n the Ka band on the Olympus satellte by the CSTS (Centro Stud sulle Telecomuncazon Spazal) Insttute (Mlan, Italy), on behalf of the Italan Space Agency. The Carrer/Nose Power (C/N 0 ) factor s montored at each staton and, on the bass of ts values, dfferent bt and codng rates are appled to lmt the BER below a chosen threshold of 10 7. Sx dfferent fadng classes are defned n ths case correspondng to φ(t) {0.0, 0.15625, 0.3125, 0.625, 0.8333, 1.0}. As sad n secton IV, the bandwdth reducton at the queue s computed as θ real (t) = φ θ (t),. As θ real (t) s just the bandwdth aval-
MARCHESE et al.: TITOLO... 7 Fg. 11. Heterogeneous aggregaton (VoIP and vdeo) wth fadng: φ(t) fadng levels over tme [25]. Fg. 13. Heterogeneous aggregaton (VoIP and vdeo) wth fadng: vdeo PLP. Fg. 12. Heterogeneous aggregaton (VoIP and vdeo) wth fadng: VoIP PLP. Fg. 14. Heterogeneous aggregaton (VoIP and vdeo) wth fadng: utlzaton of the avalable bandwdth. able for data traffc, θ (t) has to be tuned over tme n order to mantan the requred QoS. As RCBC and PID set automatcally the bandwdth under the measured losses, they do not need to know the exact value of φ(t). On the other hand, EqB follows the bandwdth computaton rule n (7), not dstngushng between the two traffc classes (n short, not usng the extenson to the multclass case to smplfy the measure of m (k) and σ (k)), but ncludng the bandwdth reducton factor φ(t). The appled mnmum bandwdth estmaton s therefore: θ mn (k + 1) = ( 1 φ(k + 1) ) m (k) + 2ln(P LP ) ln(2π) σ (k) (12) Havng just one DVB queue, the ndex would not be necessary here, but t s mantaned for coherence wth (7). P LP s the most restrctve requrement between the two nvolved traffc 1 classes. The φ(k+1) quantty corrects the bandwdth computaton n dependence of the current nformaton rate dedcated to protecton codes. EqB, used wthout multclass extenson, s not able to tune the bandwdth n proporton to the actual presence of ongong VoIP or vdeo calls. RCBC and PID, on the other hand, automatcally dstngush the bandwdth allocaton between VoIP (requrng 10 2 of PLP) and vdeo (requrng 10 3 of PLP). Fg. 11 shows the value of φ(t) over tme. Fgs 12 and 13 show the acheved PLP for each servce under dfferent traffc loads; Fg. 14 deals wth the overall utlzaton of the avalable bandwdth. The relaton between CAC and bandwdth allocaton s clear. RCBC reveals to be more precse n tunng the bandwdth; the PLP curves are almost flat and safely stay a lttle below the correspondng targets. Ths has clearly an mpact on CAC, thus obtanng the hghest levels of utlzaton for all the consdered traffc. The PLP performance of PID and EqB s good, but fals to stay below the targets wth the hghest loads. Ths can be solved by settng the up value to about 0.17 (value valdated by other smulaton results, not reported here), but ths also affects the utlzaton. The maxmum achevable utlzaton s around 0.8 (surprsngly, obtaned by all the technques) under a load of 5. It s fnally worth notng that the EqB has been appled wthout any optmzaton of the OH sze. Anyway, even f the performance of EqB can be mproved by usng the concept of domnant tme scale [11] and by applyng the multclass case so takng dstngushed measures of m,j (k) and σ,j (k) dependng on traffc class j, EqB seems not so effcent n the bandwdth allocaton/cac context presented n the paper. VIII. CONCLUSION AND FUTURE WORK The paper proposes a new approach for bandwdth allocaton and Call Admsson Control over the DVB Return Channel of a satellte system. It s sutable for the optmzaton of the bandwdth under Qualty of Servce constrants. The problem s addressed by outlnng the archtectures and the protocols of the enttes nvolved n the decson process. In more detal, the RCST control plane proposed n ths paper acts at layer 2 and ncludes a Resource Manager (RM) structured nto Decson Makers (DMs). DMs mplement vrtual queues, whch are mrrors of real traffc buffers and represent the core of the control scheme presented n ths paper. DMs compute the mnmum bandwdth to support a gven qualty and communcate ths bandwdth value to the RM, whch allocates bandwdth to traffc buffers at the data plane. Bandwdth control s joned to a Call Admsson Control acton so composng an overall control scheme wthn RCSTs. The results reported n the paper, val-
8 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.?, NO.?, MONTH? dated under dfferent condtons ncludng a real fadng trace, show promsng performance and open the door to future evolutons such as the delay and jtter control and the mplementaton of the control scheme wthn a real network node [26]. REFERENCES [1] ETSI TR 101 790. Dgtal Vdeo Broadcastng (DVB); Interacton Channel for Satellte Dstrbuton Systems; Gudelnes for the use of EN 301 790. Techncal Report. V1.2.1, (2003). [2] ETSI EN 301 790. Dgtal Vdeo Broadcastng (DVB); Interacton Channel for Satellte Dstrbuton Systems. V1.5.1, (2009). [3] ETSI TS 102 462. Satellte Earth Statons and Systems (SES), Broadband Satellte Multmeda (BSM); Broadband Satellte Multmeda Servces and Archtectures: QoS Functonal Archtecture. Techncal Specfcaton, V0.4.2, (2006). [4] M. Marchese, M. Mongell, On-lne Bandwdth Control for Qualty of Servce Mappng over Satellte Independent Servce Access Ponts, Computer Networks, vol. 50, no. 12, Aug. 2006, pp. 2089-2111. [5] Applcatons layer QoS n DVB-RCS systems, Research Project, European Space Agency, http://telecom.esa.nt/telecom/www/object/ndex.cfm?fobjectd=27802, Oct. 2007. [6] European Satellte Communcatons Network of Excellence (SatNEx) I&II, http://www.satnex.org/, and SatNExIII http://telecom.esa.nt/telecom/www/object/ndex.cfm?fobjectd=30263. [7] H. Sknnemoen, H., A. Vermesan, A. Luoras, G. Adams, X. Lobao, VoIP over DVB-RCS wth QoS and Bandwdth on Demand, IEEE Wreless Commun. Mag., vol. 12, Oct. 2005, pp. 46-53. [8] Adaptve Resource Management and Optmzaton n Satellte Networks, G. Gambene, Ed., Sprnger Verlag, Berln, Dec. 2006. [9] S. Braccal, R. Fantacc, T. Pecorella, L. Chsc, M. A. Vazquez Castro, Proactve vs. Reactve DVB-RCS Termnal Usng ACM Technques, Proc. IEEE Internatonal Conference on Communcatons 2008 (ICC 2008), Bejng, Chna, May 2008, pp. 1881-1885. [10] M. Luglo, C. Roset, F. Zampognaro, Improvng Performance of TCP- Based Applcatons over DVB-RCS Lnks, Proc. IEEE Internatonal Conference on Communcatons 2009 (ICC 2009), Dresden, 14-18 June 2009, pp. 1-6. [11] S. Georgoulas, P. Trmntzos, G. Pavlou, K. Ho, Measurement-based Admsson Control for Real-tme Traffc n IP Dfferentated Servces Networks, Proceedngs of IEEE Internatonal Conference on Telecommuncatons (ICT 2005), Capetown, South Afrca, May 2005, pp. 1-5. [12] S. Georgoulas, P. Trmntzos, G. Pavlou, K. Ho, An Integrated Bandwdth Allocaton and Admsson Control Framework for the Support of Heterogeneous Real-tme Traffc n Class-based IP Networks, Computer Communcatons, vol. 31, no. 1, Elsever, Jan. 2008, pp. 129-152. [13] M. Marchese, M. Mongell, Measurement-based Computaton of Generalzed Equvalent Bandwdth for Loss Constrants, IEEE Commun. Letters, vol. 11, no. 12, Dec. 2007, pp. 1007-1009. [14] M. Grossglauser, D. Tse A Framework for Robust Measurement-Based Admsson Control, IEEE/ACM Trans. on Networkng, vol. 7, no. 3, June 1999, pp. 293-309. [15] R. Guérn, H. Ahmad, M. Naghshneh, Equvalent Capacty and ts Applcaton to Bandwdth Allocaton n Hgh-Speed Networks, IEEE J. Select. Areas Commun., vol. 9, no. 7, Sept. 1991, pp. 968 981. [16] J. F. Whtehead, Global Packet Dynamc Resource Allocaton for Wreless Networks, US Patent 6295285, Sept. 2001. [17] M. C. Chuah, Method for Admttng New Connectons Based on Measured Quanttes n a Multple Access System for Communcatons Networks, US Patent 6377548, Apr. 2002. [18] ETSI TR 101 985. Satellte Earth Statons and Systems (SES), Broadband Satellte Multmeda (BSM); IP over Satellte. Techncal Report. V0.2.0, (2002). [19] ESA Project Feasblty study of a moble Ku-band termnal, Contract No. 15593/01/NL/DS, Fnal Report, avalable at http://telecom.esa.nt/feasblty_ndsatcom. [20] H. Sknnemoen, P. T. Thompson, "Overvew of DVB-RCS+M and ts Development," Int. J. of Sat. Commun. and Net., vol. 28, Gen. 2010, pp. 119-135. [21] ETSI TR 102 768. Dgtal Vdeo Broadcastng (DVB); Interacton channel for Satellte Dstrbuton Systems; Gudelnes for the use of EN 301 790 n moble scenaros. Techncal Report. V1.1.1, (2009). [22] N. Celandron, F. Davol, E. Ferro, A. Gotta, "Vdeo Streamng Trasfer n a Smart Satellte Moble Envronments," Int. J. of Dg. Multmed. Broadcast., 2009, 12p., artcle ID 369216, http://www.hndaw.com/journals/jdmb/2009/369216.htm. [23] F. Davol, M. Marchese, M. Mongell, Bandwdth Adaptaton for Vertcal QoS Mappng n Protocol Stacks for Wreless Lnks, Proc. IEEE Global Communcaton Conference 2009 (Globecom 2009), Honolulu, Hawa, 30 Nov.- 4 Dec. 2009, pp. 1-6. [24] N. J. Keon, G. Anandalngam, Optmal Prcng for Multple Servces n Telecommuncatons Networks Offerng Qualty-of-Servce Guarantees, IEEE/ACM Trans. Networkng, vol. 11, no. 1, Feb. 2003, pp. 60-80. [25] N. Celandron, F. Davol, E. Ferro, Statc and Dynamc Resource Allocaton n a Multservce Satellte Network wth Fadng, Internat. J. of Satellte Commun. and Networkng, vol. 21, no. 4-5, July-Oct. 2003, pp. 469-487. [26] Emulator for an ETSI BSM-Complant SI-SAP Interface, Research Project - ARTES 5 Workplan 2007, European Space Agency, http://telecom.esa.nt/telecom/www/object/ndex.cfm?fobjectd=29973, June 2009. Maro Marchese (S 94-M 97-SM 04) was born n Genoa, Italy n 1967. He got hs "Laurea" degree cum laude at the Unversty of Genoa, Italy n 1992 and the Qualfcaton as Professonal Engneer n Aprl 1992. He obtaned hs Ph.D. (Italan "Dottorato d Rcerca") degree n "Telecommuncatons" at the Unversty of Genoa n 1996. From 1999 to 2004, he worked wth the Italan Consortum of Telecommuncatons (CNIT), by the Unversty of Genoa Research Unt, where he was Head of Research. From February 2005 he has been Assocate Professor at the Unversty of Genoa, Department of Communcaton, Computer and Systems Scence (DIST). He s the founder and stll the techncal responsble of CNIT/DIST Satellte Communcatons and Networkng Laboratory (SCNL) by the Unversty of Genoa. He chared the IEEE Satellte and Space Communcatons Techncal Commttee from 2006 to 2008. He s author and co-author of about 200 scentfc works, ncludng nternatonal magaznes, nternatonal conferences and book chapters and of the book "Qualty of Servce over Heterogeneous Networks", John Wley & Sons, Chchester, 2007. Hs man research actvty concerns: Satellte and Rado Networks, Transport Layer over Satellte and Wreless Networks, Qualty of Servce and Data Transport over Heterogeneous Networks, Emulaton and Smulaton of Telecommuncaton Networks and Satellte components. Maurzo Mongell got hs "Laurea" degree cum laude at the Unversty of Genoa, Italy n 2000 and the Qualfcaton as Professonal Engneer n Aprl 2002. He obtaned hs Ph.D. (Italan "Dottorato d Rcerca") degree n "Electronc and Computer Engneerng" at the Unversty of Genoa n 2004. Hs Ph. D. was funded by Selex Communcatons S.p.A (Selex). He worked for both Selex and the Italan Consortum of Telecommuncatons (CNIT), by the Unversty of Genoa Research Unt from 2000 to 2009. He s now a member of the research staff of the Telecommuncaton Networkng Research Group by the Unversty of Genoa, wth a CNIT research poston. Hs man research actvty concerns: QoS archtectures, resource allocaton and optmzaton algorthms for telecommuncaton systems.