Published in IET Communications Received on 9th September 2009 Revised on 17th April 2010 doi: 10.1049/iet-com.2009.0587
ISSN 1751-8628
Adaptive dynamic bandwidth allocation
algorithm with sorting report messages
for Ethernet passive optical network
W.-P. Chen
1
W.-F. Wang
2
W.-S. Hwang
1
1Department of Electrical Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung 807, Taiwan 2Department of Information Engineering, National Yunlin University of Science and Technology, Yunlin 64002, Taiwan
E-mail: wwf@yuntech.edu.tw
Abstract: Broadband access networks using passive optical network (PON) technology can extend the transmission distance and increase the transmission capacity of carrier networks. One PON solution, the Ethernet passive optical network (EPON), can provide huge bandwidth capacity, low cost, simple architecture and easy maintenance. Therefore EPON becomes a promising candidate for future last-mile solutions. To prevent data collision and ensure efficient transmission, EPON must employ a media access control (MAC) protocol to allocate the shared resource of a common upstream transmission medium. This article proposes a novel DBA algorithm that sorts all REPORT messages by the request length at the next transmission cycle to fully utilise the idle time between cycles as long as at least one optical network unit (ONU) requests a long enough transmission window. Alternately, when no grant length is long enough, then some of ONUs’ requests are laid out together in the idle period to utilise the otherwise wasted idle time. Event-driven simulations show that Sort-DBA can significantly improve the network performance in terms of packet delay, average queue length and throughput, as compared with the well known IPACT, DBA2 and scheduling control scheme DBA algorithms.
1
Introduction
Internet applications such as electronic commerce, multimedia file sharing, voice over Internet protocol (VoIP) and storage area networks (SANs) have led to an enormous increase in bandwidth requirements. This tremendous growth of Internet traffic has aggravated the lack of access networks. The ‘last-mile’ between local area networks (LANs) and metropolitan area networks (MANs) remains the main bottleneck. The most widely deployed broadband solutions today are digital subscriber line (DSL), community antenna television (CATV) and cable modem (CM)-based networks. These networks, however, do not provide enough bandwidth to support the growing demand. In addition, these networks have a limitation that the distance of any xDSL subscriber to a central office (CO) must be less than 1800 feet because of signal distortion. Therefore a next-generation broadband access
network must provide not only increased bandwidth at low cost to end users, but also must provide service for long-distance access networks.
Passive optical networks (PONs) [1– 14] have aroused interest from both industry and academia as a feasible and cost-effective solution. A great deal of effort has gone into developing and standardising various PON technologies. EPONs [7 – 14] represent the convergence of low-cost Ethernet equipment (e.g. switches and network interface cards) and optical fibre architecture. Considering that more than 90% of data traffic originates from and terminates in Ethernet LANs in this moment, EPONs appear to be a natural choice for future last-mile solutions. This technology is currently being developed and standardised by the IEEE 802.3ah task force [14] in the hope of significantly improving broadband service while minimising equipment, operation and maintenance costs.
An EPON is a point-to-multipoint (P2MP) network consisting of one optical line terminal (OLT) and multiple optical network units (ONUs). The OLT broadcasts to all ONUs simultaneously in the downstream direction, whereas in the upstream direction, a single optical fibre channel is shared by all ONUs. To avoid data collision, a scheduling algorithm is needed to prevent simultaneous transmissions. The well-known media access control (MAC) protocol CSMA/CD is a standard for Ethernet LANs [8], but does not represent a good choice for EPONs. Since the OLT will receive all data packets transmitted by the ONUs and discard those involved in collisions, each ONU would require an additional receiver operating at the upstream wavelength and a carrier sensing circuit. This solution would greatly increase the network cost. Besides, its bandwidth utilisation is extremely low because the collision packets make too many data retransmissions especially under heavy traffic load conditions. Time division multiplexing (TDM) technology is a popular alternative for EPONs. Each ONU is assigned a timeslot for data transmission in each cycle and can only transmit data in the allocated window. It can be either static or dynamic, depending on the arbitration mechanism implemented by the OLT. Kramer and Pesavento[9]studied the performance of EPON under TDM using a fixed bandwidth assignment algorithm. Although this scheme is easy to implement and performs well under heavy load conditions, it cannot handle statistical multiplexing between ONUs. Static schemes based on TDM are also very inefficient because of the bursty nature of access network traffic.
To cope with this problem, Kramer et al.[10]proposed a polling-based scheme called interleaved polling with adaptive cycle time (IPACT). This algorithm achieves good performance by combining limited service with a maximum transmission window defined over 2 ms polling cycles. However, the idle time issue is not effectively resolved and the fact that IPACT allocates bandwidth based on a single ONU REPORT is not globally optimised. Owing to the bursty nature of Ethernet traffic and encapsulation of Ethernet packets (i.e. packet fragmentation is prohibited according to IEEE 802.3[14]), some ONUs may have less traffic to transmit, whereas other ONUs may have more traffic to transmit and need more bandwidth in each transmission cycle.
To address the issue, Luo et al. [11] proposed a DBA scheme called limited sharing with traffic prediction (LSTP) that predicted the arriving traffic during the waiting time and maintained a portion of the bandwidth for delivery. However, the prediction scheme has the behaviour of a bursty traffic, so some bandwidth may be wasted because the scheme cannot estimate accurately the real traffic load demand for all ONUs at the next transmission cycle. Assi et al. [12] proposed a DBA algorithm, which utilises the excessive bandwidth of lightly loaded ONUs to carry some of the bandwidth demand of
heavily loaded ONUs in each transmission cycle, thus improving the performance of the limited allocation scheme. In addition, also addressing the idle time issue, the authors proposed an early allocation mechanism, called DBA2, which schedules a lightly loaded ONU without delay, whereas it schedules heavily loaded ONUs after the OLT receives all REPORT messages and performs computation for bandwidth allocation. However, the DBA2 algorithm improves the idle period only under light or medium traffic loads. Moreover, most of the ONUs may have a bandwidth demand larger than the minimum guaranteed bandwidth under high traffic loads, so the GATE message cannot be transmitted early to the ONU for idle time compensation. In 2006, Zheng [13] proposed a bandwidth allocation called new scheduling control that uses a tracker value to address the idle time problem under high traffic loads. Although this algorithm improves the DBA2 idle time issue under heavy load conditions, it still wastes bandwidth under heavy load conditions because of the redundant overheads of the processing time of the tracker and the regular REPORT messages. Besides, it has an unfairness issue such that when the previous ONUs are operating under a light load, the following ONUs can share the remainder bandwidth, but the previous ONUs cannot allot the remainder bandwidth if the following ONUs are operating under a light load.
From the above, to improve bandwidth utilisation so as to address the idle time issue under medium or heavy traffic load conditions, we propose a new DBA algorithm called the Sort-DBA algorithm in which the transmission order of all grant data are allowed by the REPORT length at the next cycle time. The goal of this transmission scheme is to minimise the idle time under any traffic load conditions. In fact, the Sort-DBA algorithm can completely eliminate the idle time between cycles as long as at least one ONU has a sufficiently long data transmission. Even when most of the ONUs are operating under a medium load, with the waste of some guard-band distance, the new algorithm can still achieve good results with regard to eliminating idle time. Moreover, this paper also proposes a queue management method, which reduces the unnecessary overhead of REPORT messages when the ONUs are operating under heavy traffic, thereby achieving higher bandwidth utilisation under heavy traffic in EPON systems. It will be shown in our simulation experiment that Sort-DBA has better bandwidth utilisation, lower delay and lower queue length than IPACT [9], DBA2 [11] and the scheduling control scheme DBA[12]on EPON systems.
The remainder of the article is organised as follows. Section 2 summarises the basic EPON architecture. The new Sort-DBA algorithm supporting EPON is detailed in Section 3. Section 4 discusses the assumptions behind our system simulation and describes the results of simulation comparing Sort-DBA to IPACT, DBA2 and scheduling control DBA algorithms. Section 5 concludes with a few remarks.
IET Commun., 2010, Vol. 4, Iss. 18, pp. 2230 – 2239 2231
2
EPON architecture
There are several multipoint topologies suitable for an EPON: bus, ring, tree and tree-and-branch. The most popular choice is based on a star topology. As shown in
Fig. 1, it consists of one optical line terminal, a 1:N passive star splitter (and combiner) and multiple ONUs. The number of ONUs (N ) is typically between 4 and 64, but networks with N ¼ 128 have also been fabricated. The OLT resides in a CO that connects the access network to a metro core network or wide area network (WAN). The OLT is connected to a passive star splitter by a single optical fibre. The passive splitter is generally located far from the CO, but close to the subscriber premises. An ONU may be located at a curb or building, or even on the subscriber premises, and is connected to the passive splitter by a short, dedicated optical fibre. The distance between the OLT and an ONU typically ranges between 10 and 20 km. Presumed to be compatible with the IEEE 802.3 standard, all data is encapsulated in Ethernet packets for transmission. The fragmentation of Ethernet packets in the transmission window is not allowed. All transmissions occur between the OLT and the ONUs.
In the downstream direction, the OLT connects all ONUs as a point-to-multipoint (P2MP) architecture. It broadcasts Ethernet frames to all ONUs simultaneously through the 1:N splitter on a single wavelength (e.g. 1550 nm). This behaviour is similar to that of a shared media network. In the downstream direction, the Ethernet standards fit the EPON architecture perfectly: packets broadcast by the OLT are given a MAC address, so they will be extracted only at the intended destination (that is, an ONU). In the upstream direction, an EPON is a multipoint-to-point (MP2P) network. All ONUs transmit their data to the OLT on a common wavelength (e.g. 1310 nm) through the 1:N passive combiner. Since the ONUs share the upstream
transmission medium, an EPON must efficiently allocate uplink access and avoid data collisions. A MAC-based mechanism is generally chosen for this purpose.
The multipoint control protocol (MPCP) [15] has been widely used to implement DBA in EPONs. MPCP is a signalling protocol currently being developed and standardised by the IEEE 802.3ah task force. At the moment, MPCP does not specify a particular bandwidth allocation algorithm. Rather, it provides an effective control mechanism, which facilitates the implementation of bandwidth allocation algorithms. MPCP has two operation modes: normal and auto-discovery. In the normal mode, MPCP relies on GATE and REPORT Ethernet control messages to allocate bandwidth. A GATE message is used by the OLT to allocate a transmission window. REPORT messages are used by ONUs to communicate their local conditions to the OLT. In its auto-discovery mode, the protocol relies on three control messages: REGISTER, REGISTER_REQUEST and REGISTER_ACK. These are used to discover and register a newly connected ONU and to collect related information such as round-trip time (RTT) and its MAC address.
3
Sort-DBA algorithm for EPON
systems
In a global polling-based DBA algorithm, each ONU sends normally a REPORT message to the OLT after transmitting its grant data. The purpose of these messages is to request bandwidth for the next cycle time. Once the OLT has received all REPORT messages, it calculates the appropriate bandwidth allocation and broadcasts GATE messages to all ONUs while the upstream bandwidth is idle between the time of the last REPORT message (RN)
is received and the time the GATE message (G1) is sent to
ONU1. FromFig. 2, the idle time is given by
Tidle= Tdba+ RTT + TONU (1)
where RTT is the round-trip time from ONU to OLT, Tdba
is the processing time of the DBA algorithm and TONUis the
processing time of the ONU (on receiving a GATE message).
In terms of bandwidth utilisation, the idle time defined above is undoubtedly wasted. To address this problem, we propose a novel DBA algorithm by sorting all REPORT messages with the request length to give a transmission order. This DBA algorithm can be divided into two cases to eliminate the idle time according to the bandwidth demands of all ONUs. The two cases are described as follows
3.1 At least one REPORT is long enough
In fact, the idle time issue of a global polling-based DBA algorithm occurs because the OLT must wait for the last REPORT message before executing the DBA algorithm and then transmitting GATE messages to ONU1. Thus, if
the last REPORT message (RN) can be transmitted before
the last grant data (LN) and if the LN is long enough to
compensate for the idle period of formula (1), then the idle period can be completely eliminated. To ensure low delay, other grant data must be transmitted before the REPORT
messages and the transmission order must be allowed by the grant length. Hence, to minimise the idle time, the minimum guaranteed bandwidth of the last granted bandwidth per cycle (Lmin) should be at least
Lmin= Ru(Tidle− Tg) (2)
where Ruis the total upstream bandwidth of the fibre and Tg
is the guard-band time between two neighbouring packets. For example, inFig. 3, the last REPORT message (RN21)
received by the OLT before the last grant data (LN21) at cycle
time I can use up some of the idle time that would normally occur before cycle time (I+ 1). During a DBA calculation time, the OLT sorts all of the request lengths in order to find the size of request lengths R1, RN, R4, . . . and R3;
furthermore, R1is long enough to compensate for the idle
period at cycle time (I+ 1). The OLT transmits a series of GATE messages (GN, G4, . . . , G3, G1) to all of the
ONUs. Since the length of the granted bandwidth (LN21)
is long enough at cycle time I, the OLT is sufficient time to execute the DBA algorithm and transmit GATE messages to all of ONUs under the idle period. In this way, the idle time problem can be neatly solved as long as at least one ONU has a sufficiently long data transmission.
If all of the ONUs are under heavy traffic load, the maximum cycle time (Tmax) is given by formula (3), where
Figure 2 Idle time issue in an EPON polling scheme
Figure 3 Example for the Sort-DBA algorithm under high traffic load conditions
IET Commun., 2010, Vol. 4, Iss. 18, pp. 2230 – 2239 2233
under most load conditions than other DBA algorithms. To fully utilise the idle time between transmission cycles, it calculates the maximum transmission window (unlike IPACT) for each ONU demand and adjusts the order of ONUs so that the longest transmission comes last. At medium and light loads, when none of the ONU transmissions is long enough to completely use up the idle time on its own, a compensation scheme is developed so that one or many grant data are laid out together after last grant data to fill the idle space. The presented simulation results clearly show that Sort-DBA allows significant improvement in bandwidth utilisation by reduction of the idle period. Furthermore, the proposed algorithm further enhances performance by using a queue management scheme to reduce the overhead of REPORT messages when the ONU queues are long. IPACT algorithms (with fixed service and limited service), DBA2 (an improved IPACT scheme) and the scheduling control scheme (an improved DBA2 scheme) were compared to our DBA algorithm in an event-driven EPON network simulation. The proposed DBA algorithm significantly achieves better performance than other DBA algorithms, with average maximum throughput per ONU and system efficiency of around 62.15 Mbps and 99.44%, respectively.
Our future work will investigate how to support a QoS scheme in a Sort-DBA algorithm. Secondly, we are extending our simulations to cases involving asymmetric traffic loads.
6
Acknowledgment
The authors would like to thank The National Science Council (NSC) of Taiwan for supporting this research under project number NSC 98-2221-E-151-037.
7
References
[1] KIM K.S.: ‘On the evolution of PON-based FTTH solutions’, Inf. Sci., 2003, 149, (1 – 3), pp. 21 – 30
[2] LUNG B.: ‘PON architecture “future proofs” FTTH’, Lightw. Mag., 1999, 16, (10), pp. 104 – 107
[3] ANGELOPOULOS J.D.,BOUKIS G.C.,VENIERIS I.S.: ‘Delay priorities enhance utilisation of ATM PON access systems’, Comput. Commun., 1997, 20, (11), pp. 937 – 949
[4] ANGELOPOULOS J.D., LELIGOU H.C., ARGYRIOU T., ZONTOS S., RINGOOT E.,VAN CAENEGEM T.: ‘Efficient transport of packets with QoS in an FSAN-aligned GPON’, IEEE Commun. Mag., 2004, 42, (2), pp. 92 – 98
[5] MAIER G.,MARTINELL M.,PATTAVINA A.,SALVADORI E.: ‘Design and cost performance of the multistage WDM-PON access networks’, IEEE Lightw. Technol., 2000, 18, (2), pp. 125 – 143
[6] SKUBIC B.,CHEN J.,AHMED J.,WOSINSKA L.,MUKHERJEE B.: ‘A comparison of dynamic bandwidth allocation for EPON, GPON, and next-generation TDM PON’, IEEE Commun. Mag., 2009, 47, (3), pp. S40 – S48
[7] KRAMER G., MUKHERJEE B., PESAVENTO G.: ‘Ethernet PON (ePON): design and analysis of an optical access network’, Photonic Netw. Commun., 2001, 3, (3), pp. 307 – 319
[8] CHAE C.J., WONG E., TUCKER R.S.: ‘Optical CSMA/CD media access scheme for ethernet over passive optical network’, IEEE Photonics Technol. Lett., 2002, 14, (5), pp. 711 – 713
[9] KRAMER G., PESAVENTO G.: ‘Ethernet passive optical network (EPON): building a next-generation optical access network’, IEEE Commun. Mag., 2002, 40, (2), pp. 66 – 73
[10] KRAMER G.,MUKHERJEE B.,PESAVENTO G.: ‘IPACT: a dynamic protocol for an Ethernet PON (EPON)’, IEEE Commun. Mag., 2002, 4, (1), pp. 74 – 80
[11] LUO Y., ANSARI N.: ‘Limited sharing with traffic prediction for dynamic bandwidth allocation and QoS provisioning over EPONs’, J. Opt. Netw., 2005, 4, (9), pp. 561 – 572
[12] ASSI C., YE Y., DIXIT S., ALI M.A.: ‘Dynamic bandwidth allocation for quality of service over Ethernet PONs’, IEEE J. Sel. Area. Commun., 2003, 21, (9), pp. 1467 – 1477 [13] ZHENG J.: ‘Efficient bandwidth allocation algorithm for Ethernet passive optical networks’, IEE Proc. Commun., 2006, 153, (3), pp. 464 – 468
[14] IEEE 802.3ah task force, http://www.ieee802.org/3/ efm
[15] WAN packet size distribution available from http:// www.nlanr.net/NA/Learn/packetsizes.html
[16] ZHU Y.,MA M.: ‘IPACT with grant estimation (IPACT-GE) scheme for Ethernet passive optical networks’, J. Lightw. Technol., 2008, 26, (14), pp. 2055 – 2063
