Resource Management in Satellite Networks part 24 doc

214 Tommaso Pecorella, Giada Mennuti Simulations have evaluated the bandwidth loss, measured as the number of slots per frame allocated to an RCST but not used to transmit data, and the QoS performance of two DBA techniques based on a Receding Horizon Controller and a Smith Predictor Controller (RHC and SPC, respectively) versus three fixed (Fix n in Figure 7.3, where n is the number of slots/frame assigned to a single RCST) allocation schemes with different bit-rate values [13]. A self-similar traffic with Hurst parameter H = 0.8 has been used to feed each RCST. Figure 7.3 reports the bandwidth loss distribution (a) and the delay for the Expedited Forwarding (EF), Assured Forwarding (AF), and Best Effort (BE) DiffServ [14]-[17] traffic classes [(b), (c) and (d), respectively]. It can be observed that the bandwidth loss is greatly reduced by using DBA techniques, whereas the overall QoS of the traffic classes is acceptable, in particular by using the RHC scheme. However, despite the great advantages of DBA techniques, both in terms of QoS satisfaction and efficient resource utilization, some new problems arise when they are applied. In particular: • “Greedy” traffic flows can compromise the whole satellite system’s QoS. • Compatibility between different control techniques and bandwidth request methods should be validated. • Security issues on the signaling channel should be analyzed in order to prevent denial of service attacks based on fake bandwidth reservations. Those issues should be investigated and carefully addressed before using DBA techniques in any actual system. 7.3.2 Dynamic bandwidth de-allocation Several approaches for bandwidth and handover management have been studied in the recent literature in the case of mobile satellite systems. Publications in this area investigate only bandwidth allocation and the intra- satellite handover management. In reference [18], an advanced bandwidth management strategy is proposed and evaluated, allowing for bandwidth allo- cation/deallocation and a novel inter-satellite handover management scheme, tailored for multimedia LEO satellite networks with satellite diversity. The main mechanism is based on bandwidth de-allocation. According to the proposed scheme, capacity reservation requests for handover calls are removed from the queues when the capacity that they strive to reserve is unlikely to be used. Simulations confirmed the usefulness of bandwidth de-allocation mechanism. Other details of this scheme have been already discussed in sub-Section 6.4. 7.3.3 Dynamic bandwidth allocation with cross-layer issues Some examples of cross-layer DBA schemes are here briefly discussed, limiting the description to recent works [19]-[25]. An overview of cross-layer approaches Chapter 7: DYNAMIC BANDWIDTH ALLOCATION 215 Fig. 7.3: DBA in GEO satellite systems. 216 Tommaso Pecorella, Giada Mennuti in this context can be found in [26],[27]. The first example considered concerns the presence of mixed Guaranteed Bandwidth (GB) real-time traffic and Best Effort (BE) traffic. GB is subject to CAC, which is exerted independently by local controllers, situated at the Earth stations (i.e., RCSTs), within an amount of bandwidth that is assigned to these stations over a certain allocation time interval. The bandwidth allocation is the responsibility of a master station and can be done periodically or on-demand. Best-effort traffic is represented by an inelastic model (i.e., the congestion control mechanism of TCP is not explicitly taken into account), used to compute the loss probability of cells (Asynchronous Transfer Mode, ATM or DVB), stored in the Earth stations’ buffers; this traffic utilizes the bandwidth that remains available after serving the GB class. The cross-layer interaction stems from the fact that a fade countermeasure, based on bit and coding rate adaptation, is used at the physical layer, whose influence on the bandwidth allocation is accounted for by “redundancy coefficients” [representing the inverse of the ratios of the Information Bit Rate (IBR) in the specific channel condition to the one in clear sky]. Various methods have been considered for bandwidth allocation, and the overall structure has been evaluated in the presence of real fading traces [19],[20]. Figure 7.4 [20] represents call blocking, call dropping (due to a temporary lack of bandwidth) and cell loss probabilities for three different allocation strategies: (i) cross-layer Optimized Centralized (OC, where the bandwidth is allocated on demand by the master station, which solves a centralized optimization problem); (ii) cross-layer Optimized Proportional (OP, where optimal allocation requests are computed locally by the Earth stations and then passed to the master, which re-scales them and distributes the bandwidth proportionally); (iii) Simple Proportional (SP, based on offered load, with no cross-layer dynamic allocation). The reported results refer to a 10,000 s simulation, with 10 Earth stations, 5 of which experience different fading conditions, whereas 5 operate in clear sky. In these graphs, the probabilities for each point in time are computed by averaging over all stations in the system, and over a time window of 1,000 s. The fading is dynamically variable, according to real traces. The advantage of the cross-layer allocations lies in maintaining blocking probability values below a given threshold (5% in the specific case), while minimizing the call dropping and the BE traffic cell loss probabilities in the stations’ buffers. The second example deals with DBA in the presence of only inelastic packet traffic with two stations, whose traffic loads periodically alternate between a lower and a higher value. Figure 7.5 [22] illustrates the convergence properties of a gradient descent technique, based on Infinitesimal Perturbation Analysis (IPA) [21]-[23]. Station 2 is in clear sky, whereas station 1 also experiences fading variations, besides those in traffic load. The bandwidth allocation provided by the IPA gradient estimation, based only on on-line measurements, is capable to face both dynamic effects in order to minimize Chapter 7: DYNAMIC BANDWIDTH ALLOCATION 217 Fig. 7.4: Call blocking and dropping (left) and cell loss (right) probabilities. These graphs are reproduced from “Adaptive Cross-layer Bandwidth Allocation in a Rain-faded Satellite Environment”, N. Celandroni, F. Davoli, E. Ferro, A. Gotta, International Journal of Communication Systems, Vol. 19. No. 5, pp. 509–530, June 2006. c 2006. Copyright John Wiley & Sons Limited. Reproduced with permission. Fig. 7.5: IPA gradient descent allocation, under traffic load and fade changes. See reference [22]. Copyright c 2006 IEEE. the overall loss volume. The problem considered in [21],[22] is a pure parametric optimization. In order to avoid transient periods in the convergence of the on-line gradient descent technique, a different point of view can be adopted [23] where open- loop feedback control strategies (i.e., stemming from a functional optimization approach) are approximated by means of neural networks. Finally, a DBA cross-layer optimization, aiming at achieving the “best” compromise between the TCP goodput maximization and fairness, has been treated in [24],[25], in a GEO bent-pipe satellite scenario. The numerical details of the example shown here are the same as in [25], with a combination 218 Tommaso Pecorella, Giada Mennuti of long-lived TCP NewReno connections, sharing various bottleneck links, determined by 10 different fading classes (stemming from different source- destination pairs), under the Hotbird 6 link budget [28] and real fading traces. The “instantaneous” goodput is determined by the dynamic bandwidth and redundancy allocation, which aims at counteracting fading effects and achiev- ing a compromise between maximizing the total goodput and maintaining fairness among connections. Figure 7.6 [25] shows the behavior as a function of time of the overall goodput (the points are the results of a moving average over a 10 s window) for two classes operating under different fading conditions (note that the carrier power-to-noise spectral density ratio, C /N 0 , for the source-destination pairs is also shown). The used strategies attempt to maximize the total goodput and to maintain fairness in different ways (see [25] for a description of these strategies): • The “merge” strategy is the best choice between two alternative methods (“tradeoff” and “range”, respectively) that establish a balance between goodput and fairness; • The “proportionally fair” technique maximizes the sum of the logarithms of the individual goodputs, so as to attain a Nash Bargaining Solution (NBS); • The “BER threshold” strategy simply adjusts the redundancy to keep always BER below a given limit, and assigns the bandwidths proportion- ally to the redundancy and the number of connections of each class (no cross-layer action). The advantages of the cross-layer strategies, shown in detail in [25], are not only in terms of goodput, but also in terms of fairness. 7.3.4 Joint timeslot optimization and fair dynamic bandwidth allocation in a system employing adaptive coding In [29], an enhanced and multi-beam DVB-RCS system is addressed, consid- ering both Adaptive Coding (AC) and dynamic framing. AC arises when the transmission is severely affected by channel conditions (as in the Ka band). In order to keep the link active, framing design must be flexible enough to adapt in time and frequency, to allow for the use of different carriers (this technique is also known as Dynamic Resource Allocation, DRA) and/or different protection-levels of channel coding (AC). The problem of optimal framing has been already addressed in the literature. For example, in [30] a method is presented for optimal super-frame pattern design for the DVB-RCS MF-TDMA return link, so that the system data throughput is maximized. The authors formulate the design problem as a non-linear combinatorial optimization problem. However, the developed method considers static framing and, therefore, it is not extensible to Ka band Chapter 7: DYNAMIC BANDWIDTH ALLOCATION 219 Fig. 7.6: Merge, Proportionally Fair and BER Threshold (thr = 10 −6 )strategies. A class in fading (a); a class in clear sky (b). See reference [25]. Copyright c 2006 IEEE. 220 Tommaso Pecorella, Giada Mennuti transmissions, where adaptive physical layer is used and adaptive framing is more appropriate. In [31], bandwidth segmentation on a super-frame basis is also presented, but slots are of fixed duration within the different types of carriers. Current systems typically make use of fixed framing for its simplicity. Instead, the DVB-RCS terminal considered in [29] is assumed to have trans- mission capabilities sufficiently agile to realize multimedia communications under adverse channel conditions. This is achieved by using the dynamic-slot MF-TDMA feature of the standard, integrated with AC that is adapted on a super-frame basis. This strategy segments the total bandwidth into different types of carriers, according to user traffic demands and weather conditions. It is assumed that slots can be assigned to users with different coding rate on the same carrier, which leads to coexisting slots of different size, which is called adaptive framing, due to AC. The total bandwidth is segmented into several carriers that can be of different bandwidths and the slots contained in the frames can be of various durations, according to the chosen coding rate; users can be granted slots of different durations on different carriers (sequentially). Differently from other studies, bandwidth is segmented not only in the presence of different traffic types, but also assuming realistic dynamic weather conditions, to which coding rate is adapted. Capacity is allocated giving priority to heavy rain-affected users, then considering less affected ones, and ending with clear sky users, while there is still bandwidth available. The major issue to keep into account concerns the limits of capacity that can be allocated, due to adaptive framing. In this study, the time dimension is partitioned into super-frames, a super-frame into frames and frames into slots. The super-frame length is 26.5 ms and seven different coding rates (1/3, 2/5, 1/2, 2/3, 3/4, 4/5, 6/7) are considered; the modulation is QPSK. Regarding the frequency dimension, it is assumed that the total bandwidth can be dynamically segmented, from super-frame to super-frame, and that up to four different carrier types can be used in a super-frame: 540 kHz (carrier type I), 270 kHz (carrier type II), 135 kHz (carrier type III) and 67.5 kHz (carrier type IV). The roll-off factor is 0.35, providing symbol rates of 400, 200, 100 and 50 kbaud, respectively. The number and type of active carriers is adapted to the traffic requests and the needs of the users, which vary according to channel conditions. The transmitted packet can be an ATM cell or an MPEG packet (for numerical evaluations we will only refer to ATM cells). With AC, the length of the slots transmitting such fixed-length packet becomes variable and, therefore, the number of slots contained by a given type of carrier becomes variable, as well. Not all the users are necessarily always active. Active users are divided into categories, according to both their symbol energy to noise-plus-interference spectral density ratio, E s /N o,tot and traffic characteristics. Traffic is assumed to be uniform and the considered classes are: Constant Bit Rate (CBR), Variable Bit Rate (VBR), and BE. For simplicity and without loss of gen- erality, one user is assumed to ask only for one of these traffic classes, so that Chapter 7: DYNAMIC BANDWIDTH ALLOCATION 221 each user will have allocated slots of a given fixed length on the carrier type corresponding to its E s /N o,tot [29]. Traffic demands are queued according to the type of DVB-RCS capacity request, which can be CRA, RBDC, and VBDC. Capacity requests are prioritized: CRA has the highest priority and VBDC the lowest. CBR traffic is assigned to CRA as a whole, whereas VBR traffic is assigned to CRA and RBDC. Similarly, BE traffic is also divided between RBDC and VBDC. The number of carriers of each type is computed at every super-frame, given priority to the users affected by rain. Assuming a given E s /N o,tot for the user and some given requests for the current super-frame, a closed-form estimation of the number of carriers required per carrier type is computed in terms of an estimation of the number of slots as follows: n C n i (s)=n C,CBR n i (s)+n C,V BR n i (s)=N C n i (s)  ( r C CBR +r C VBR ) T s η i L(η i )  , i =1, 2, , N AC (7.1) n R n i (s)=n R,V BR n i (s)+n R,BE n i (s)=N R n i (s)  ( r R VBR +r R BE ) T s η i L(η i )  , i =1, 2, , N AC (7.2) n V n i (s)=n Vol n i (s)+n V,BE n i (s)=N V n i (s)  V +r V BE T s η i L(η i )  , i =1, 2, , N AC (7.3) where s is an index making reference to the super-frame, which consists of 10 frames and lasts 265 ms, n X,Y n i (s) is the number of requested slots corresponding to capacity request type X (X = C, R or V, which correspond to CRA, RBDC or VBDC requests, respectively) of traffic class type Y (Y = CBR, VBR or BE) requiring spectral efficiency η i , n X n i (s) is the total number of requested slots corresponding to capacity request type X, N X n i (s)isthe number of users requesting capacity type X, T s is the duration of a super-frame in seconds. N AC is the number of possible coding rates, and r X Y is the bit-rate requested by traffic class Y that is mapped to request type X. V is a possibly additional amount of bits requested as volume (instead of bit-rate), which results in n Vol n i (s) slots, and L(η i ) is the length in bits of the packet. The number of carriers of each type is estimated from the total number of slots needed according to (7.1)-(7.3). The fragmentation of the bandwidth into carriers is performed, starting from the heavy rain-affected users down to the clear sky ones, while there is still bandwidth available. With all these assumptions, a key result has been obtained in [29] by applying cross-layer design for DVB-RCS with AC. The user satisfaction strongly depends on the distribution of users relative to the spatial distribution of channel conditions. As a conclusion, smarter scheduling policies should be designed, taking into 222 Tommaso Pecorella, Giada Mennuti account this effect in order to design fairer bandwidth allocation schemes. In what follows, a smarter scheduling policy is proposed, based on a joint optimization accounting for both a fair bandwidth distribution among users and channel conditions and timeslot duration. Proposed framework Recall that in DVB-RCS, the TBTP (composed of several frames (F) of duration T F ) is updated and transmitted every super-frame. If BW is the total system bandwidth, then the scheduler is in charge of solving an allocation problem for each BW × T F block (note that it can also be applied to the whole super-frame). BW is divided into different carrier types to serve different users, accounting for different Service Level Agreements (SLAs), location and terminal equipment. The problem to be faced consists in multiplexing N users into C carriers of BW i bandwidth that transmit into a frame of T F seconds. An ETSI specification [32] imposes a number of constraints to the problem, namely: • The total transmission capacity (i.e., carriers) in the satellite beam is divided in areas. • The symbol rate and slot timing must be the same for all carriers in one area. Coding rates are not necessarily the same. • A given RCST belongs to one (and only one) area and can use only one carrier at a given time. Hence, it is possible to simplify the problem creating sub-problems, one for each group of carriers of the same type (see Figure 7.7, on the left [9]). It is meaningful to consider that the RCSTs in one area, while transmitting in a common carrier type, use the same transmission rate. Note that the DVB-RCS standard defines an adaptive-coding physical (PHY) layer with several possible coding rates, so the mapping of users to areas is basically defined by the quality of the link (channel conditions). As before, the minimum transmission unit (a layer-2, MAC, packet) can be an ATM cell (53 bytes) or a Moving Picture Experts Group (MPEG) container (188 bytes). The following analysis is related to the case of ATM cells. Following the previous discussion, the aim here is to obtain TBTP reduced signaling for frame description (excessive signaling in the Frame Composition Table, FCT, entails a reduction in bandwidth efficiency). A timeslot with com- mon duration for all areas is imposed ( 1 ), allowing a very simple assignment 1 Note that fixing a timeslot duration common to all areas introduces some unused bandwidth that depends on both the timeslot duration and the packet length (ATM cell in our case). However, once a given RCST has been assigned to a certain timeslot, it can change its transmission rate inside the timeslot without affecting the transmission timing of the other RCSTs. This argumentation validates the robustness of the solution proposed. Chapter 7: DYNAMIC BANDWIDTH ALLOCATION 223 Fig. 7.7: Scheduling (bandwidth allocation) problem. See reference [9]. Copyright c 2006 IEEE. procedure (after having known the number of timeslots per area): from left to right and from top to bottom (according to the reading order). Regarding signaling issues, this is translated into a simple FCT, since it indicates the common timeslot type (which is described in the Time Composition Table, TCT) and how many times it is repeated in the carrier. On the basis of the area rate, one or more ATM cells can be transmitted in a single timeslot. A possible timeslot and ATM cell assignment is shown in Figure 7.7, on the right [9]. The problem of how to assign timeslots to areas and ATM cells to RCSTs is discussed later, after introducing the scheduling hierarchy concept [32]. Scheduling hierarchy The general scheduling problem (which may involve thousands or more RCSTs) may be complex to solve. Therefore, it seems reasonable to reduce it to some smaller problems by imposing some known structure (that can also facilitate signaling). This is an idea similar to that proposed in [33] (particularly in centralized optimization algorithms). According to [32], some minimum resources are guaranteed to the service providers. Since the relative RCSTs for each service provider can be distributed over different areas, in [32] the scheduling hierarchy presents the segment concept, i.e., a grouping of . sharing various bottleneck links, determined by 10 different fading classes (stemming from different source- destination pairs), under the Hotbird 6 link budget [28] and real fading traces. The “instantaneous”. novel inter -satellite handover management scheme, tailored for multimedia LEO satellite networks with satellite diversity. The main mechanism is based on bandwidth de-allocation. According to. different fading conditions, whereas 5 operate in clear sky. In these graphs, the probabilities for each point in time are computed by averaging over all stations in the system, and over a time window