Self-Control Cyclic Access with Time Division - A MAC Proposal for The HFC System

Transcription

1 Sef-Contro Cycic Access with Time Division - A MAC Proposa for The HFC System S.M. Jiang, Danny H.K. Tsang, Samue T. Chanson Hong Kong University of Science & Technoogy Cear Water Bay, Kowoon, Hong Kong ABSTRACT: The IEEE 82.4 standard committee is currenty working on a project to find a cost-effective means of providing access to integrated networks for peope to enjoy mutimedia programs and to work at home. An advanced system based on the CATV system caed Hybrid Fiber Coax is being studied. Since some properties of the HFC system precude the possibiity of directy using existing medium access contro protocos for its data ink ayer, a MAC scheme based on time division is discussed in this paper.. Introduction As ATM technoogy progresses, researchers are ooking for a cost-effective means of providing home access to integrated networks so that peope can enjoy mutimedia programs and work at home. Due to the arge geographica coverage of existing CATV systems, an advanced system based on CATV caed Hybrid Fiber Coax (HFC) is being studied by the IEEE 82.4 Project Committee []. This system wi ikey use ATM technoogy to provide mutipe services to home such as teephone, TV and information retrieva. In the HFC system, the downstream band and upstream band operate asymmetricay at different frequencies, and a node has ony one ink to each bus. Therefore, the HFC system may be abstracted into a ogica dua-branch-bus based network consisting of a distributing hub (caed headend in this paper) and the fiber nodes (or simpy nodes) as shown in Figure 3. In the downstream direction (from the headend to the nodes) broadcast mode is used, and the headend is the ony transmitter. This makes contro of the downstream band simpe. In the upstream direction (from the nodes to the headend), the nodes share the upstream band for transmission, and contention access mode is used. To aow each node to have equa opportunity to share the upstream band, an efficient medium access contro (MAC) protoco is necessary. This protoco has to be cost-effective and efficient for ong distance communication, supporting a wide range of transmission rates as we as both anaog and digita broadband services [2]. Unfortunatey, these properties and requirements make it difficut to use existing MAC protocos directy for the foowing reasons: ) The ring-based protocos such as Token-Ring [5], FDDI [5], ATMR [3] and Orwe [4] are not suitabe for the duabranch-bus topoogy required by the HFC system. On the other hand, both CSMA/CD and Token-Bus [5] are suitabe for the HFC topoogy. However, CSMA/CD cannot guarantee bounded deay and is ony suitabe for reativey short distances. The weakness of Token-Bus is that it requires the users to remain on-ine to be incuded in the ogica ring which needs to be reconfigured whenever users are added or removed from the network. 2) It is difficut to use MAC protocos based on the sotted media approach ike DQDB [5] and CRMA [6]. These protocos often require a node to have two inks to each bus for data insertion. Moreover, DQDB needs two headends and CRMA requires a foded bus whie there is ony one ogica headend and the topoogy is required to be dua-branch-bus in the HFC system. Athough the two ogica buses in the HFC system can be ooped to form a foded bus ike CRMA, the outbound bus (from the headend to the nodes) is reserved for transmission, and the inbound bus (from the nodes to the headend) for receiving in the foded-bus CRMA. This is opposite to the HFC system where the downstream band is for the users to receive data and upstream band for the users to transmit data. With regards to the specia position of the headend in the HFC system, a MAC scheme based on centraized contro is a natura choice. One such scheme is the traditiona poing system such as Ro-ca Poing [8]. However, these systems have been designed for ow or medium speed data appications under the assumption that the propagation deay is negigibe compared to the transmission time. This assumption is not true in high speed networks [7]. In this paper, we wi discuss a new MAC scheme caed Sef-Contro Cyce Access with Time Division (SCCA/TD). The rest of the paper is organized as foows. The principe of SCCA/TD is introduced in Section 2, and simuation studies on this prototype are give in Section 3. Finay Section 4 concudes the paper. 2. Introduction to SCCA/TD SCCA/TD is a reservation based MAC scheme which aows the nodes to share the upstream bandwidth. A node normay needs to make requests for bandwidth through the upstream band. The headend wi respond to the requesting nodes through the downstream band. The upstream bandwidth is divided into equa time segments caed cyces. A cyce is divided again into fractions which are assigned to the nodes according to their requests. As shown in Figure, the same epoch of the cyce is repeated periodicay as ong as the requests of the nodes are unchanged. That is, the quota and position of the bandwidth aocation to a given node in each cyce are not changed. In this way, a node can access the medium simpy knowing the reative position of its assignment in the cyce. The granted bandwidth can be used by the nodes for transmitting data or submitting requests. Cyce k Cyce k+ Cyce k+2 3 i 3 i 3 i 2.. Cyce Structure i assignment to node i : cyce ength Figure The principe of SCCA/TD

2 A cyce is divided into the static reservation segment (SRS) and the dynamic reservation segment (DRS) as shown in Figure 2. SRS gives each node a basic bandwidth assignment periodicay in each cyce. A node can use the SRS assignment to submit requests or transmit sma amounts of data. DRS, on the other hand, provides bandwidth dynamicay to the nodes according to their requests. The advantage of using SRS is that it provides the nodes quick access to the medium. However, if a node has nothing to transmit, its assignment wi be wasted. Requiring severa nodes to share the same assignment aternatey can improve efficiency. When a node has a arge amount of data to transmit, it needs to get the requested bandwidth from DRS. To this end, the node submits a request to the headend carried by a specia packet containing the node's identifier, transmission speed, request type and other reated parameters. The headend wi respond with a specia confirmation packet. cyce: L SRS DRS SRS (Static Reservation Segment) is staticay reserved to each node DRS (Dynamic Reservation Segment) is aocated to nodes according to the node's requests L / L2: Length of SRS / DRS, = L + L2 L2 Figure 2 The cyce structure There are two types of requests that a node can submit to the headend: Gross Bandwidth Request (GBR) and Cycic Bandwidth Request (CBR). With GBR, a node informs the headend of the tota amount of bandwidth it needs, and the headend serves the request in consecutive cyces if the request cannot be satisfied within a singe cyce. In this case, the requesting node wi get the tota required bandwidth in mutipe cyces, each providing ony a fraction of the tota required bandwidth. With CBR, a node informs the headend of the bandwidth it requires in each cyce. If the request is granted, the node gets the same quota of bandwidth in the same reative ocation in each subsequent cyce unti the end of transmission. When the node finishes transmission, it has to send a message to the headend to reinquish its DRS reservation. The first type of request is suitabe for bursty traffic such as data whie the second type is suitabe for periodic traffic such as voice and video. For both SRS and DRS, the nodes can adopt the sefcontro cycic access mechanism to use the bandwidth after receiving confirmation without additiona contro from the headend. That is, the headend notifies the requesting node the starting time of its bandwidth assignment, the quota of bandwidth that the node can use in each cyce and the cyce ength, i.e., the period in which the node can use its assignment unti either the expiration of its grant (for GBR request) or the end of transmission (for CBR request). This scheme can be impemented with timers or counters for fixed ength MAC protoco data units Scheduing Schemes Coisions burden networks with extra work for coision detection and resoution, and wastes bandwidth. This is especiay undesirabe in high-speed networks. Therefore, SCCA/TD is designed to be coision-free. The scheduing done by the headend is the key to coision avoidance. The bandwidth aocation mechanism has to maintain fairness for the nodes to share bandwidth whie avoiding coision to improve network utiization. There are many aternatives for the aocation poicy. A simpe scheme woud be for the headend to coect the requests and serve them one by one in round robin. This scheme is used in the simuation studies beow. In the foowing, we discuss two coision avoidance approaches. The foowing parameters are used in the discussion: n - tota number of nodes in the network (see Figure 3); m - tota number of nodes simutaneousy sharing the same cyce; - cyce ength; d(i,j) - distance between nodes i and j; d - distance between nodes and n, i.e., maximum propagation deay; µ - measured in terms of the tota time used for transmission within a cyce over the cyce ength. (for simpicity,, d(i,j) and d are a measured in time units) a) The Efficient Approach (EA) headend Tb 2... i... m... n upstream band bus current cyce downstream band bus Te / Tb Te() next cyce Tb() d(,2) Te(2) Te() no transmission Tb(2) d(2,3) Te(3) Te(m) Tb(3)... Tb(m) d(,m) Tb() Tb / Te: Beginning / End of the cyce Tb(i) / Te(i): Beginning / End of assignment to node i Figure 3 Diagram of cyce division with EA For this approach, we suppose that the signa propagation in the upstream band bus is unidirectiona towards the headend, and the sequence of bandwidth assignment to the nodes in each cyce is in accordance with the sequence of their physica positions in the network. In this case, the cyce aways begins with the assignment to the node cosest to the headend. The propagation deay between adjacent nodes is taken into account to maximize which can go up to. This can be expained by Figure 3, where the cyce is fuy shared by m active nodes. Notice that an active node i+ can advance its transmission by d(i,i+) ahead of the end of transmission of node i due to propagation deay. The tota time advanced within a cyce is equa to d(i,i+) = d(,m) (i =, 2,.. m- ). When the ast active node m finishes transmission, time d(,m) is required to aow its traffic pass through a the downstream nodes before the beginning of the next cyce. This time is compensated by the tota time advanced by a the nodes. Therefore, EA can give high independent of network size and cyce ength. However, if no bandwidth in a suitabe position in the cyce is avaiabe to a new requesting node, the headend has to shift the positions of some existing assignments in order to accept new requests. For exampe, as shown in Figure 4, if nodes 3, 4 and 8 become active, there is no free segment suitabe to them even though there are severa free segments in the cyce. By shifting the bandwidth assignments for nodes, 2 and 7 to the eft (occupying some of the unassigned space), some spaces are made avaiabe for the new requesting nodes in the right position. In this case, the headend needs to send messages to the affected nodes informing them their new

3 bandwidth assignments in the cyce. Shifting bandwidth assignments can provide more fexibe and efficient bandwidth aocation at the expense of compexity. Cyce division: Inactive nodes: ➂ ➃ ➇ ➉ Active nodes: ❶ ❷ ❺ ❻ ❼ ❾ occupied Figure 4 Assignment in a cyce versus the physica position of the nodes A simpe impementation of EA caed quasi-ea () is that when aocating an assignment in the cyce for a node, the position of the assignment in the cyce is determined by the node's physica position in the network to avoid irreguar cyce divisions. b) The Simpe Approach () With this approach, the reative physica positions of the nodes in the network are not taken into account. The maximum distance (i.e., d) is used as the distance between any pair of nodes. In this case, when an active node finishes its assignment in a cyce, the next active node has to wait d time units before starting its transmission. µ = - m *, where = d /, is the bandwidth wastage per active node due to the gap paced between adjacent assignments for avoiding coision. As shown in Figure 6, is inversey proportiona to the number of active nodes for a given cyce ength and network size. The maximum is bounded by -. Therefore, this approach is suitabe for sma size networks ony because it is easy to get a sma with ong cyces in this case. 3. Simuation Studies As mentioned in Section 2., there are two types of requests that a node can make, i.e., GBR and CBR. With CBR, a node can get the guaranteed bandwidth in every cyce if its CBR request is confirmed. This is simiar to GBR when a arge GBR request is divided into mutipe fragments served by severa consecutive cyces. Therefore, our simuation is concerned with GBR requests ony. 3. Simuation Mode The simuation mode consists of 25 nodes with equa distance between any two neighboring nodes (s). The network size (NS) is equa to 24 * s (assuming the distance between node and the headend is ). The traffic mode of each node is a bursty puse train [9] as shown in Figure 5. The deay between two consecutive bursts and the number of ces in a burst foow respectivey the exponentia and geometric distributions. The network capacity is set to Mbps corresponding to OC-3, which resuts in the ce ength (CL: time to transmit an ATM ce) of µs. The static reservation for each node is set to s. The ength of the static reservation segment in a cyce, L, is equa to 25 * s. The mean ce arriva rate from a node (MCR) can be computed by BL / (DB + BL * IC). The ce transmission rate given by the network is equa to / CL, which is shared by 25 nodes so free that each node gets / (25 * CL). The offered oad (OL) is defined as MCR / [ / (25 * CL)] = 25 * CL / MCR. Given CL, OL and BL, DB = (25 / OL - ) * BL * CL. The main performance index to be investigated is µ. The request response time is not measured since a node can continue to make requests even after the previous ones have not been honored; and the headend may honor severa requests for a given node at the same time. Tabe ists the main parameter vaues used by the different simuations (NS = network size, BS = buffer size, = cyce ength, OL = offered oad, BL = burst ength, PV = parameter to be varied). The defaut M setting is (L2 - I) for the with I = d, and L2 / 25 for the scheme. DB burst BL DB: mean deay between two burrsts (exponentia distribution) BL: mean number of ces in a burst (geometric distribution with probabiity =. / [BL + ]) IC: interva between two ces (constant equa to a ce ength) Figure IC Figure 5 Burst puse train traffic mode NS (km) BS (ce) (ce) OL BL (ce) Fig PV. 3 Fig PV PV 3 Fig.8 3 PV 3.75 PV Fig.9 3 PV 3.75 PV Fig. 3 PV 3.75 PV Fig. PV 25 PV. 3 Fig.2 PV 25 PV. PV Tabe Parameter setting for the different simuations 3.2 Simuation Resuts This subsection is organized as foows: Figures 6 to investigate the effects of some parameter settings (e.g., request threshod, offered oad and buffer size) to the performance of and s; and Figure compares the performance between and. Athough shows better performance than in many cases as shown beow, the main advantages of are its simpicity and appicabiity to wireess networks. The simuation resuts are obtained using the simuation package BONes DESIGNER V 2.6 from Ata Group. Each point on the graphs took about 3 hours to simuate on a SPARC-5 machine. Figure 6 shows the effect of request threshod (T) to (µ) for and. It is observed that T has very itte effect on µ. For, a reasonabe cyce setting (e.g., CL = 3) can produce a good µ. For, the cyce ength is 5 ces and M is set to 5, 9, 25 (average setting) and 45 (maximum setting) ces respectivey. Note that can amost achieve the same µ as, but the number of simutaneous active nodes in is imited to or 2 by the maximum M setting (i.e., M = L2 - I) whie that in is equa to N (e.g., N = 25 in this simuation mode). Figure 7 depicts (µ) versus offered oad (OL) for different cyce engths (). µ at T is the theoretica maximum vaue (TMµ) for a given network size and cyce ength according to the anayses in Section 2.2. Note that it is easier for traffic to fi a shorter cyce (e.g., = ) rather than a onger one. However, increasing µ with

4 increasing OL in with a short cyce is imited by the TMµ which is smaer than in. It is possibe to have a high µ in since the TMµ is equa to. In addition, different cyces in ead to the same µ in the case of heavy oad since µ given by is not infuenced by the cyce setting but mainy depends on the traffic pattern of the nodes ike the EA scheme mentioned in Section = =3 = request threshod (ce) a) µ vs, M = - 5 ces M = 5 M = 9 M = 25 M = request threshod (ce) b) µ vs M, = 5 ces Figure 6 Effect of request threshod to Figures 8 ~ 9 show the performance (bandwidth utiization, mean response time and mean ce oss rate) for different buffer sizes (BS) and different burst engths (BL). Generay, when the buffer is arge enough, the infuence of BL becomes weak in, but this is not the case in. It is because M is set to the maximum in so that a node can get a arger quota in each cyce. After finishing the quota, the node needs to wait for a ong time to get another quota. During this period, the arge buffer can be used to smooth the traffic. However, adopts the average setting so that a node can usuay get a quota in each cyce. This makes sensitive to traffic burstiness. Therefore, with the average setting is more suitabe for ess bursty traffic = =3 = offered oad (OL) = =3 = T offered oad (OL) Figure 7 Bandwidth utiization versus offered oad T mean response time (ms) mean response time (ms) BL= BL=3 BL=5 75 BL= BL=3 BL=5 75 Figure 8 Bandwidth utiization versus buffer size BL= BL=3 BL= BL= BL=3 BL=5 75 Figure 9 Mean response time versus buffer size Figure 9 shows the mean response time (MRT) as a function of buffer size. MRT is the eapsed time between the ce arriva to the buffer and its departure. Different from, MTR in is sensitive to BL due to the reason mentioned above. Figure presents the mean ce oss rate (MCLR) cacuated in terms of tota ost ces over tota generated ces. When the buffer size is sma, MCLR for is smaer than that for because the bandwidth aocation in the cyce in with the average setting is more suitabe than that in with the maximum setting for the traffic accumuated in a sma buffer. The effect is reversed in the case of arge buffers due to the reason mentioned above. Figure shows (µ) versus network size (NS) with cyce engths of 3 and 5 ces. _t and EA_t are respectivey the theoretica maximum µ for and EA respectivey (refer to Section 2.2 for the cacuation). It is cear that µ in decreases with increasing NS in the case of = 3. In the case of = 5, µ first increases with NS up to NS = 52 then decreases. This is due to the mismatch between the assignment in the cyces and the traffic oad from

5 the nodes. Since ong distance increases the response deay because of the propagation deay, this gives some time for the nodes to accumuate traffic. Simiar observation is aso found in. However, this increase is imited by the theoretica maximum vaue. Anyway, is ess infuenced by NS than so is more suitabe for metropoitan area of networks. mean ce oss rate mean ce oss rate BL= BL=3 BL= BL= BL=3 BL=5 75 Figure Mean ce oss rate versus buffer size _t EA_t a) = 3 ces 65 network size (km) _t EA_t b) = 5 ces network size (km) Figure Bandwidth utiization versus network size The fairness with which the nodes share the upstream bandwidth depends on the service discipine used by the scheduing agorithm at the headend. In the simuation, the nodes are served one by one in round-robin rather than FIFO discipine since the atter wi ead to unfairness caused by the propagation deay between the nodes and the headend. Figure 2 pots the mean access interva for each node, which is the mean interva between two consecutive accesses to the media for a node. As shown in Figure 2.a, behaves better than since use the average setting and uses the maximum setting. As iustrated in Figure 2.b, performs very we in the case of burst ength equa to. A better service discipine is needed to ensure fairness among a nodes for BL >. mean access interva (µs) mean access interva (µs) node index a) =5 ces, NS=65 km, BL=3 ces 5 5. Discussion BL= ces BL=3 ces node index b), =3 ces, NS=26 km Figure 2 Fairness issue on mean access interva in and A MAC scheme based on time division for the HFC system, SCCA/TD, was discussed in this paper. The reservation scheme with sef-contro-cycic-access mechanism, which can be easiy impemented with timers, aows SCCA/TD to provide the bounded deay and throughput. This scheme requires ess contro from the network and permits the nodes to be passivey inked to the buses which enhance network robustness. The non-sotted media based framework can accommodate different transmission rates and support different types of MAC protoco data units. Athough the priority mechanism has not been addressed, it can be easiy impemented since the aocation poicy is centray performed by the headend. The static reservation segment (SRS) in each cyce can provide quick access to the medium. The scheduing agorithm is the key in SCCA/TD. Two schemes were discussed in the paper: the simpe one () is suitabe for sma size networks and the efficient one (EA) suitabe for a wide range of network sizes. More efficient schemes to reaize EA require further studies. References [] J. W. Eng, J. F. Moenauer, IEEE Project 82.4: Standards for Digita Convergence, IEEE Communications Magazine, May 995 Vo. 33 No. 5, pp2-23 [2] Mario P. Vecchi, Broadband Networks and Services: Architecture and Contro, IEEE Communications Magazine, August 995 Vo.33 NO.8, pp24-32 [3] Kazuo Imai,Tadashi Ito,Hideki Kasahara et a., "ATMR: Asynchronous transfer mode ring protoco", Computer Networks and ISDN Systems 26 (994) [4] J.L. Adams, "Orwe", Computer Networks and ISDN Systems 26 (994) [5] James Martin, "Loca Area Networks, Architectures and Impementations", 2nd Edition, PTR PRENTICE HALL, 994 [6] E.A. Zurfuh, R.D. Cideciyan et a, "The IBM Zurich Research Laboratory's.3 Gb/s LAN / MAN Prototype", Computer Networks and ISDN Systems 26 (993) [7] Bandua W.Abeysundara, "High-speed oca area network and their performance:a survey", ACM computing surveys,vo.23,no.2,june 99,p22-65 [[8] Mischa Schwartz, "Teecommunication Networks, Protocos, Modeing and Anaysis", Adison-Wesey pubishing Company, 988 [9] Ata Group of Cadence Design Systmes (Inc), BONeS DESIGNER Core Library Reference, V 3.