Dynamic Bandwidth Allocation Schemes to Improve ...

1 downloads 0 Views 673KB Size Report
xDSL(x Digital Subscriber Line) and cable modem technolo- gies provide relatively ... In Interleaved Polling with Adaptive Cycle Time. (IPACT) [6], they also ...
IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

1

Dynamic Bandwidth Allocation Schemes to Improve Utilization under Non-Uniform Traffic in Ethernet Passive Optical Networks Kyuho Son, Hyungkeun Ryu, Student Member, IEEE and Song Chong, Member, IEEE

Abstract— Ethernet Passive Optical Networks (EPONs) are an emerging access network technology that provides a low cost solution for fiber-to-the-home (FTTH) and fiber-to-the-business (FTTB). In this paper, conventional bandwidth allocation schemes in EPON are shown to suffer from poor utilization under nonuniform traffic conditions, particularly as the number of ONUs, guard time and round-trip time increase. To resolve this problem, we propose a new scheme that allocates a timeslot intelligently by considering other ONUs’ queue occupancy, instead of strictly enforcing a maximum timeslot size. The analysis and simulation results show that the proposed scheme can provide significantly higher utilization than conventional schemes and can support max-min fairness under non-uniform traffic conditions. Index Terms— access network, Ethernet Passive Optical Network (EPON), bandwidth allocation scheme, non-uniform traffic, maximum utilization, max-min fairness, performance analysis.

I. I NTRODUCTION Recently, the capacity of backbone networks has been increased dramatically. However, in access networks, dial-up, xDSL(x Digital Subscriber Line) and cable modem technologies provide relatively small increases in data transfer. With the rapid growth of the number of Internet users and multimedia services, access networks have become bottlenecks. A passive optical network (PON), which is a point-to-multipoint optical network composed solely of passive elements, is regarded as the technology most likely to offer a solution to this problem, [1], [2] and is considered to be the best way to support emerging services such as high-quality digital television broadcast services and online education multicast services. In particular, Ethernet PON (EPON) is the best candidate for a next generation optical access network because Ethernet is cheap, simple, scalable and popular, and is regarded as an economical method for deploying fiber-to-the-X (FTTX) [3]. Fig. 1 shows a typical topology of EPON. An optical line terminal (OLT), located at a local exchange, is connected to multiple optical network units (ONUs) through passive elements, splitter and coupler. In downlink transmission from OLT to ONUs (using 1550nm wavelength), a packet is broadcasted automatically to all ONUs through an optical splitter and each ONU filters the received packet according to its destination address. In uplink transmission from ONUs to OLT (using 1310nm wavelength), a packet from all ONUs can only reach the OLT, not other ONUs. As a result, ONUs Manuscript received January xx, 2004; revised November xx, 2004. This work was supported in part by University IT Research Center Project and Electronics and Telecommunications Research Institute, Korea.

must share a single uplink optical fiber trunk. To prevent data from being corrupted due to multiple ONUs transmitting at the same time, a medium access control protocol is required. Therefore, a Multi-Point Control Protocol (MPCP) is being developed by the IEEE 802.3ah task force to arbitrate between transmissions by ONUs to avoid collisions [4]. The MPCP is operated principally by two control messages, GATE and REPORT. Each ONU informs the OLT of the queue occupancy by REPORT to help the OLT make a decision about bandwidth allocation. The OLT assigns the transmission timeslot by GATE. In EPON, one of the most important issues is how to share uplink resources efficiently and fairly (i.e., the bandwidth allocation problem), and much research is currently being done by experts on this topic. Kramer et al. [5] suggested a fixed service scheme that always grants the maximum timeslot size to each ONU. They determine the optimal maximum timeslot size by using a utilization-delay optimization approach. In Interleaved Polling with Adaptive Cycle Time (IPACT) [6], they also proposed a well-designed bandwidth allocation framework that minimizes the unused timeslot by interleaving polling messages. IPACT can reduce the polling delay while maintaining high link utilization, and can be used with a variety of allocation schemes: fixed, gated, limited, constant/linear credit service, etc. Kramer et al. compared the above schemes through simulations and concluded that neither of the discussed service schemes is better than the limited service scheme. In [7], the authors introduced the concept of threshold report to achieve higher bandwidth efficiency. In addition to research on the bandwidth allocation problem, there is much research being done on differentiating quality of services (or classes of services) in EPON [8]–[11]. Previous research mostly has performed simulations under the assumption of uniform traffic conditions. In this paper, we show that not only do conventional bandwidth allocation schemes suffer from utilization degradation under non-uniform traffic conditions, but also that the degradation can be severe when the number of ONUs, guard time and round-trip time increase. We note that this problem can be resolved if an intelligent decision is made about bandwidth allocation without maximum timeslot restriction, which decision takes into account not just one, but all queues of ONUs. Our proposed scheme can provide higher utilization than conventional schemes and can support max-min fairness under non-uniform traffic conditions. The remainder of this paper is organized as follows. In

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

TABLE I D ETAILED DESCRIPTION OF GATE SCHEDULING .

Every time the OLT receives REPORT from ONUi , 1) Update SEI ← max(SEI, now + RT Ti ). start time = SEI 2) Decide the timeslot size Gi by the specific bandwidth allocation scheme. duration = Gi 3) Update SEI ← SEI + Gi + r + g. 4) Send GATE to ONUi .

section 2, we describe an EPON system model. In section 3, we present conventional and proposed bandwidth allocation schemes. In sections 4 and 5, we state the utilization analysis and various simulation results, to demonstrate that our schemes perform well. We present conclusions in section 6. II. M ODEL D ESCRIPTION Assume that an EPON has BW link capacity and consists of an OLT and ONUi , 1 ≤ i ≤ N , where N is the total number of ONUs in the EPON. Every time the OLT receives REPORT containing a request (i.e., queue length) from ONUi , it sends GATE containing the transmission start time and duration to ONUi . A detailed description and graphical explanation of this GATE scheduling1 are given in Table I and Figure 2, respectively. SEI (Scheduling End-point Indicator) is the last point that has been allocated (i.e., the earliest point that a new allocation can be started) and now is the current time. RT Ti is the round-trip time2 between the ONUi and the OLT. r is the time required to send REPORT3 , and g is the guard time including the following: the ONU’s laser turn on/off time, and the OLT’s receiver auto gain control (AGC) settling time and clock recovery time. When ONUi receives GATE from the OLT, it sends data throughout the allotted time Gi from start time, and then sends a new REPORT, which contains current queue length information Ri , at the end of data transmission. If the OLT receives a new REPORT from ONUi , it performs the GATE scheduling repeatedly. In the ordinary case of SEI ≥ now +RT Ti in Figure 2 (a), the OLT allocates all the timeslots without omission. However, in the case of SEI < now + RT Ti in Figure 2 (b), the OLT cannot receive data from ONUi directly after the current SEI even if the OLT sends GATE immediately, because the OLT must wait for a previous REPORT to send a new GATE. Due to this unallocated timeslot, utilization degradation may occur. In addition, the cycle time and the delay cannot be smaller than a certain value even if the load is extremely small. We refer to this phenomenon as the RTT effect. 1 Our GATE scheduling method is quite similar to the method proposed in IPACT [6]. 2 In EPON, the geographical distance from an OLT to ONUs can be from several km to 20km [4] and the RTT can be from several µs to 200 µs. 3 Fixed service scheme do not need REPORT message, r = 0.

2

III. BANDWIDTH A LLOCATION S CHEMES In this section, we describe conventional and proposed bandwidth allocation schemes. We consider the following to be basic bandwidth allocation schemes. 1) Fixed service scheme always grants the maximum timeslot size Gmax . 2) Gated service scheme grants as much as is requested, Gk+1 = Rik . i 3) Limited service scheme grants as much as is requested, but cannot exceed Gmax , Gk+1 = min(Rik , Gmax ). i where Gki and Rik denote the kth timeslot size for ONUi and the kth request (i.e. queue length information) from ONUi respectively. Various bandwidth allocation schemes, including the three schemes stated above, are compared in [6]. The fixed service scheme is simple, but does not take into account the amount of traffic from each ONU. This scheme is mentioned here only for comparison. The authors find that the gated service scheme can utilize the bandwidth more efficiently without any unused remainder4 , and the delay and average queue length of ONUs are much smaller than in other schemes in the midrange offered load. However, the gated service scheme is not suitable for use in EPON because the cycle time and delay may increase without bound as the load increases. Therefore, the authors conclude that neither of the schemes discussed in [6] is better than the limited service scheme. In this paper, we will compare our proposed schemes with the best-performance limited service scheme. The limited service scheme can achieve higher utilization, but may fail to fully utilize the bandwidth under non-uniform traffic conditions. Suppose that only one ONU wants to send data. The ONU can use Gmax in each cycle, where a cycle consists of Gmax and a number N of guard times and report max times. Therefore the utilization is limited to GmaxG+N (r+g) . Assuming that the number of ONUs N = 16, maximum timeslot size Gmax = 125µs, guard time g = 5µs and report time r = 0.512µs, the maximum utilization is only 58.6%. The situation becomes more serious as the number of ONU N and the guard time g increase. This is because the OLT cannot grant more than Gmax even if other ONUs do not use any resources. We can expect utilization improvements if the bandwidth allocation scheme can grant a timeslot longer than Gmax considering other ONUs’ traffic. Therefore, two proposed schemes that satisfy this condition, P1 and P2, are suggested in this paper. A. Proposed Scheme 1 (P1) Using other ONUs’ GATE information Gj for i 6= j (the recently granted timeslots of other ONUs), P1 calculates how many timeslots have already been used by other ONUs and how many timeslots remain. To perform this task, we newly define a variable Fik and set its value to the total timeslots 4 In the fixed and limited service schemes, an unused remainder of the granted timeslot may exist at the end of transmission if an ONU has more packets to send than the granted timeslot. This is because the packet size at the head of the queue may be bigger than the remainder.

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

per cycle minus the timeslots recently used by other ONUs. This variable indirectly indicates how much OLT can grant to ONUi . When Rik ≤ Gmax , P1 grants as much time as requested, k Ri , as in the limited service scheme. Otherwise, P1 guarantees at least Gmax and attempts to grant Ri as much as possible but no more than Fik . As a result, the utilization performance of P1 is at least as good as that of the limited service scheme.  k Ri if Rik ≤ Gmax k+1 Gi = k k min(Ri , max(Gmax , Fi )) otherwise (1) P(i−1) mod N where Fik = N Gmax − j=(i−N +1) mod N Gj .

where Gi is the size of the timeslot recently allotted to ONUi . The utilization will be improved by relaxing the constraint on the maximum size of timeslot Gmax . For example, when only one ONUi wants to send data, the OLT can grant N Gmax to N Gmax ONUi and the utilization will be N Gmax +N (r+g) = 95.8%. But P1 may cause a problem regarding the fairness of timeslot allocation, because it is possible that an ONU, which was originally using the N Gmax timeslot, yields not N G2max , but only Gmax to a newly arriving ONU. For better presentation, we depict this example in Fig. 3 and compare P1 with other schemes. B. Proposed Scheme 2 (P2) We now propose a new dynamic bandwidth allocation scheme (P2) to achieve both utilization improvement and maxmin fairness. P1 uses ONUs’ GATE information, while P2 uses ONUs’ REPORT information (the recently requested queue length of other ONUs). To calculate fair distribution, if an OLT sends GATEs to all ONUs once the OLT has received requests from all ONUs, the polling time will cause timeslot waste. Therefore, a scheme in which the OLT does not wait until all the requests are collected is preferable. In the time between the OLT’s receiving REPORT from ONUi and sending the corresponding GATE, the OLT will receive REPORT from a number M of other ONUs, and will not send GATE to them before it sends GATE to ONUi . Hence, the REPORT information from M ONUs is not used. However, the received REPORT information from the other ONUs (total number of ONUs N minus M ) is used, in other words, the OLT sends GATE to (N − M ) ONUs before it sends GATE to ONUi . In P2, the OLT distributes N Gmax max-min fairly among ONUs based on total N requests and sends GATE to ONUi with the amount of its own fair share, on the assumption that there is little difference between the (N − M ) information already used and the new (N −M ) information that will arrive later. P2 allocates a timeslot as follows:  if si = 1  Ri PN k+1 N G − s R max j j (2) Gi = PNj=1 otherwise  N −

where sj =

  

1

j=1

if

sj

N P

min(Rj , Rk ) ≤ N Gmax

k=1

0

otherwise

3

where Ri is the recent request from ONUi and si indicates whether or not the OLT grants as much as ONUi requests. Suppose that only one ONUi always has enough data to transmit and other ONUs do not have any data. This means that Rj = 0, sj = 1, j 6= i and si = 0, thus OLT can grant N Gmax N Gmax to ONUi and the utilization will be N Gmax +N (r+g) = 95.8%. Some features of the four bandwidth allocation schemes (fixed, limited, P1 and P2) under the non-uniform traffic are summarized in Table. II. IV. P ERFORMANCE A NALYSIS In this section, we analyze the utilization performance of the fixed, limited and proposed schemes. First, we will derive the utilization under arbitrary ONUs’ input traffic, assuming that there is no RTT effect (this does not mean that RT T = 0). Next, we will find the maximum utilization under the special non-uniform scenario in which the traffic load of ONU1 is 1 − ρ and the traffic load of ONU2∼N is ρ. Finally, we will take the RTT effect into account. A. Utilization without considering the RTT effect Let Ii and Oi respectively denote the rate of input and output at ONUi normalized to EPON link rate. We will derive output rates {Oi }i=1,2,...,N in terms of arbitrary input rates {Ii }i=1,2,...,N without considering the RTT effect. Then, the PN total link utilization will be U = k=1 Oi . 1) Fixed service scheme: The OLT always grants Gmax . Thus, the cycle time is equal to a constant c = N Gmax + N g. The output rate can be calculated as   Gmax Oi = min Ii , (3) N Gmax + N g and the output rate is the same as the input rate for all ONUi (i.e. the system is stable) when Ii ≤

Gmax , N Gmax + N g

f or all i.

(4)

2) Limited service scheme: To determine whether or not si = 1, assume that si = 1. As we mentioned earlier, si denotes whether or not the OLT grants as much as ONUi requests. If we replace input rate Ik , which is larger than Ii , by Ii or Ik0 = min(Ii , Ik ), then all ONUs will satisfy sk = 1, Ok0 = Ik0 for all k. Under the changed input rates, the cycle (r+g) time will be c0 = NP because the total input rate is N 0 1−

0

k=1

Ik

. In the limited service scheme, the allotted equal to c −Nc(r+g) 0 timeslot Gi must be smaller than Gmax , Gi = c0 Oi = c0 Ii ≤ Gmax . Therefore we can determine si as follows. ( (r+g) Ii ≤ Gmax 1 if NP N 1− I0 si = (5) k=1 k 0 otherwise PN Gmax (1−s )+N (r+g) k=1 PN i because c = The cycle time is c = 1− si I i k=1 PN PN PN k=1 Gk +N (r +g) = c k=1 si Ii +Gmax k=1 (1 − si )+

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

4

TABLE II C OMPARISONS OF FOUR BANDWIDTH ALLOCATION SCHEMES UNDER NON - UNIFORM TRAFFIC CONDITIONS . Scheme

Operation Gk+1 = Gmax i Gk+1 = min(Rik , i

Fixed Limited

Gk+1 i

=

P1

(

Information used

Gmax )

Rik



P2

where sj =

min(Gmax ,

Ri

Fik ))

otherwise

j=(i−N +1) mod N

j=1

N P

sj

O O

good

X

good

O

my own REPORT Ri and

k=1

Gj

otherwise

if

0

otherwise

(1−si )+

(6)

Gmax

N (r + g) Ii ≤ Gmax f or all i. PN 1 − k=1 Ik

(7)

3) Proposed schemes 1 & 2: The total utilizations of P1 and P2 are equal at the steady state even though Oi may be different from each other. Therefore we will consider only P2 For every i, we set si = 1 if PNin the following analysis. N Gmax min(I , I ) ≤ i k k=1 N Gmax +N (r+g) , otherwise si = 0. If there exists i such that si 6= 0, at least one ONUi is not satisfied even though OLT grants total N Gmax per cycle. In this case, the cycle time will be c = N Gmax + N (r + g). (r+g) because the total input rate is equal Otherwise, c = NP N 1−

Ik

4) Utilization comparison under uniform and non-uniform traffic conditions: To compare utilization of the fixed, limited and proposed schemes under uniform and non-uniform traffic conditions, we choose the traffic rates of ONUs as ONU1∼30 : ONU31∼32 = 1 : k where k is the non-uniformness parameter. Fig. 4 shows the utilization of the four schemes with k =1, 10, 100. Under uniform traffic conditions, that is when k=1, the four schemes have equal utilization. In fact, the proposed schemes perform almost the same as the limited service scheme in this case. Note that the utilization of the conventional schemes decreases as non-uniformness k increases. However, the proposed schemes make it possible to achieve the same utilization as under uniform traffic conditions. B. Maximum Utilization without considering the RTT effect Now, let us find the maximum utilization when the traffic load offered by the ONUs, except for ONU1 , is 0 ≤ ρ ≤ (N −1)Gmax N Gmax +N (r+g) . This can be done by putting (10) into the results of the previous subsection.

The output rate can be calculated as

  Ii 

k=1

other ONUs’ REPORT Rj

k=1

By using the cycle time and output rate, we can obtain the mean granted timeslot size for ONUi , Gi = cOi and the stable condition (si = 1, ∀ i) is given by

c−N (r+g) . c

my own REPORT Ri and

min(Rj , Rk ) ≤ N Gmax

1

N (r + g). The output rate can be calculated as  if si = 1  Ii PN (1− si I i ) Oi = Gmax k=1 otherwise  c = PN N (r+g)

Oi =

poor moderate

if si = 1

N −

 

P(i−1) mod N

PN sj R j PNj=1

N Gmax −



to

nothing

if Rik ≤ Gmax

min(Rik ,

where Fik = N Gmax − Gk+1 = i

Fairness

my own REPORT Ri

other ONUs’ GATE Gj

 

Utilization

if si = 1

N Gmax − N Gmax +N (r+g) N

N−

P

PN

k=1

k=1

sk

sk I k

otherwise

I1 = 1 − ρ and (8)

N X

N Gmax . N Gmax + N (r + g)

Ii = ρ, Ii = Oi for i 6= 1

i=2

U=

By using the cycle time and output rate, we can obtain mean granted timeslot size for ONUi , Gi = cOi and stable condition is given by

N X

N X

Oi = O1 + ρ and s1 = 0, si = 1, for i 6= 1

i=1

The maximum utilization of the fixed, limited and proposed schemes can be obtained as follows: Gmax +ρ N Gmax + N g

(11)

ULimited =

Gmax + ρN (r + g) Gmax + N (r + g)

(12)

UP roposed =

N Gmax N Gmax + N (r + g)

(13)

(9)

UF ixed =

Note that (9) does not have any condition on i. It requires only that the total input rate is smaller than the effective N Gmax capacity Cef f = N Gmax +N (r+g) . This distinguishing feature allows the proposed schemes to provide full utilization, even under non-uniform traffic conditions.

k=1

Ik ≤

(10)

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

5

C. Maximum Utilization, taking into account the RTT effect

V. S IMULATION R ESULTS

In EPON, the geographical distance from an OLT to ONUs can be from several kilometers to 20km [4] and the roundtrip time is not negligible. As seen in Fig. 5, the RTT may result in waste of timeslots if condition (14) is satisfied. This is because the OLT must wait for a previous REPORT to send a new GATE. X Gk + (N − 1) r + N g < RT Ti (14)

In this section, we present simulation results to verify the analysis and demonstrate the performance of the proposed schemes. Fig. 7 shows our simulation model using an OPNET simulator [12]. We generate the self-similar traffic by aggregating 32 pareto-distributed ON-OFF sources [13], [14]. We generate Ethernet frames exponentially distributed with mean 500bytes from 64bytes to 1518bytes. Table III shows the parameters used in this simulation.

k6=i

In this case, maximum utilization is smaller than derivations (11)∼(13); thus we need a modification. In this subsection, we will take the RTT into consideration and derive the maximum utilization as the load offered ρ by ONUs, except ONU1 . 1) Fixed service scheme: The fixed service scheme is not effected by the RTT because it does not use REPORT. Therefore, the maximum utilization is the same as in the previous derivation (11). Gmax +ρ (15) N Gmax + N g 2) Limited service scheme: Maximum utilization will be achieved when ONU1 uses Gmax per cycle. To find the minimum offered load ρth to avoid the RTT effect, we formulate the following equations from Fig. 6(a).

TABLE III S IMULATION PARAMETERS EPON Link rate (BW ) Number of ONUs (N ) Maximum timeslot (Gmax ) Control message length (r) Guard time (g) Round trip time (RTT) OLT/ONU queue size

1Gbps 8, 16, 32 125 µs 0.512µs (=64bytes) 1, 5, 10µs 50,100,200µs 300kbytes

UF ixed =

RT T = cρth + (N − 1)r + N g

(16)

c = Gmax + r + RT T

(17)

By using (16), (17), we can obtain ρth as follows.   RT T − (N − 1)r − N g , 0 ρth = max Gmax + r + RT T

(18)

When ρ ≤ ρth , the cycle time is Gmax + r + RT T . Therefore, the maximum utilization is modified in the region of ρ ≤ ρth .  Gmax   Gmax +r+RT T + ρ if ρ ≤ ρth ULimited = (19)   Gmax +ρN (r+g) if ρ > ρth Gmax +N (r+g) 3) Proposed scheme 1 & 2: Maximum utilization will be achieved when the ONUs use N Gmax per cycle. To find the minimum offered load ρth to avoid the RTT effect, we formulate the following equations from Fig. 6(b). RT T = cρth + (N − 1)r + N g

(20)

c = N Gmax + N (g + r)

(21)

By using (20), (21), we can obtain ρth as follows.   RT T − (N − 1)r − N g ρth = max , 0 N Gmax + N (r + g)

(22)

+r+RT T because When ρ ≤ ρth ,Pthe cycle time is c = N Gmax1+ρ c = N Gmax − k6=1 Gk +r+RT T = N Gmax −cρ+r+RT T . Therefore, the maximum utilization is modified in the region of ρ ≤ ρth .  N Gmax (1+ρ)   N Gmax +r+RT T if ρ ≤ ρth (23) UP roposed =  N Gmax  if ρ > ρ th N Gmax +N (r+g)

Fig. 8 depicts the maximum utilization versus network load (i.e., the offered load by the ONUs except ONU1 ). The fixed service scheme grants the fixed timeslot Gmax to every ONU even if the network load is low. As a result, the maximum utilization is very low. The limited service scheme can achieve higher maximum utilization than the fixed service scheme because ONU1 gets the chance to send as much data as the other ONUs do not use at an earlier time. However, the maximum utilization of the limited service scheme decreases as the non-uniformness increases (i.e. network load decreases), and the RTT effect prevents high maximum utilization from being achieved in the low offered load region ( ρ < ρth ). For example, maximum utilization is very low in ρ < 0.34 when RT T =100µs, g=1µs, N =16. If we increase RT T to 200µs, the situation becomes more serious. Note that the proposed schemes can always achieve almost full utilization, regardless of the network load. In addition, the maximum utilization of conventional schemes decreases as the number of ONUs and guard time increases, while the proposed schemes are almost independent of the two variables. To observe the fairness performance of the four schemes, we construct a simulation scenario as follows. ONU1 sends traffic at 300Mbps from 0sec and ONU2 starts to send 300Mbps at 10sec under 500Mbps background traffic produced by ONU3∼16 . Fig. 9 compares the throughput of ONU1 and ONU2 . Until 10sec, ONU1 sends only 60Mbps and 260Mbps, respectively, in the fixed and limited service schemes. However, both P1 and P2 are allowed to send 300Mbps all. After 10sec, the throughputs of ONU1 and ONU2 are 60Mbps in the fixed service scheme. In the limited service scheme, ONU1 and ONU2 get 180Mbps equally. ONU1 and ONU2 use 195+255=450Mbps in P1, which is 90Mbps higher than 180+180=360Mbps in the limited service scheme, but the distribution is unfair. The total utilization of P2 is the same as that of P1. However, ONU1 and ONU2 use 225+255=450Mbps fairly in P2. This means that we achieve both higher utilization and fairness using our P2 scheme.

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

TABLE IV FAIRNESS S IMULATION SCENARIO AND RESULTS

ONU1 ONU2

Arr. (sec) 0 10

Dept. (sec) ∞ ∞

Input Rate (Mbps) 300 300

Fixed 60 60

Output Rate (Mbps) Limited P1 P2 180 195 225 180 255 225

PLACE PHOTO HERE

6

Kyuho Son received his B.S. and M.S. degrees in Electrical Engineering and Computer Science from the Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea in 2002 and 2004, respectively. He is currently workging toward the Ph.D. degree in Electrical Engineering and Computer Science at the same place. His current research interests are in the areas of bandwidth allocation for ethernet passive optical networks and radio resource management for 3G/4G wireless network systems.

VI. C ONCLUSION In this paper, we point out the low utilization problem of conventional bandwidth allocation schemes under nonuniform traffic conditions. To resolve this problem, we propose new schemes that relax the maximum timeslot restriction and make an intelligent decision about bandwidth allocation using other ONUs’ information. Our proposed schemes always utilize more than 90% of bandwidth under any circumstances. P1, using previous GATE information, may fail to guarantee fairness. However, P2, using previous REPORT information, guarantees max-min fairness as well. Based on our analysis and simulation, we conclude that our proposed scheme (P2) always fully utilizes the bandwidth under non-uniform traffic conditions and guarantees max-min fairness. R EFERENCES [1] G. Pessavento and M. Kelsey, ”PONs for the Broadband Local Loop,” Lightwave, vol.16, no.10, pp.68-74, Sep. 1999. [2] B. Lund, ”PON Architecture ’futureproofs’ FTTH,” Lightwave, vol.16, no.10, pp.104-107, Sep. 1999. [3] G. Kramer, G. Pessavento, ”Ethernet Passive Optical Network (EPON): Building a Next-Generation Optical Access Network,” IEEE Communication Magazine, vol.40, issue 2, pp.66-73, Feb. 2002. [4] IEEE Draft P802.3ahTM/D1.2, ”Media Access Control Parameters, Physical Layers and Management Parameters for subscriber access networks,” Dec. 2002, http://grouper.ieee.org/groups/802/3/efm. [5] G. Kramer, B. Mukherjee, and G. Pessavento, ”Ethernet PON (ePON): Design and Analysis of an Optical Access Newtork,” Photonic Network Communications, vol.3, no.3, pp.307-319, Jul. 2001. [6] G. Kramer, B. Mukherjee, and G. Pessavento, ”IPACT: A Dynamic Protocol for an Ethernet PON (EPON),” IEEE Communication Magazine, vol.40, issue 2, pp.74-80, Feb. 2002. [7] O. Yoshihara, Y. Fujimoto, N. Oota, N. Miki, ”High Performance EPON,” IEEE 802.3ah Ethernet in the First Mile Task Force, Nov. 2001, [Online]. Available at http://grouper.ieee.org/groups/802/3/efm/public/nov01/ yoshihara 1 1101.pdf. [8] G. Kramer, B. Mukherjee, S. Dixit, Y. Ye, and R. Hirth, ”Supporting differentiated classes of service in Ethernet passive optical networks,” OSA Journal of Optical Networking, vol.1, no.8&9, pp.280-298, Aug. 2002. [9] Maode Ma, Yongqing Zhu, and Tee Hiang Cheng, ”A Bandwidth Guaranteed Polling MAC Protocol for Ethernet Passive Optical Networks,” Proceedings of IEEE INFOCOM 2003, pp.22-31, Mar. 2003. [10] C.M. Assi, Y. Ye, S. Dixit and M.A. Ali, ”Dynamic Bandwidth Allocation for Quality-of-Service Over Ethernet PONs,” IEEE Journal on Selected Areas in Communications, vol.3, no.9, pp.1467-1477, Nov. 2003. [11] Fu-Tai An, Yu-Li Hsueh, Kyeong Soo Kim, Ian M. White, and Leonid G. Kazovsky, ”A new dynamic bandwidth allocation protocol with quality of service in Ethernet-based passive optical network,” International Conference on Wireless and Optical Communication (WOC 2003), Jul. 2003. [12] OPNET Modeler 7.0, http://www.opnet.com. [13] W. Willinger, M. Taqqu, R. Sherman and D. Wilson, ”Self-similarity through high-variablity: statical analysis of Ethernet LAN traffic at the source level,” IEEE/ACM Transactions on Netwrking, vol. 5, issue 1, pp.71-86, Feb. 1997. [14] W. Leland, M. Taqqu, W. Willinger, and D. Wilson, ”On the self-similar nature of Ethernet traffic (extended version),” IEEE/ACM Transactions on Netwrking, vol. 2, issue 1, pp.1-15, Feb. 1994.

Hyung-Keun Ryu received the B.S. degree in Electronic Engineering from Kyungpook National University, Daegu, Korea and the M.S. degree in Electrical Engineering and Computer Science from PLACE Korea Advanced Institute of Science and Technology PHOTO (KAIST), Daejeon where his is currently pursuing HERE the Ph.D. degree. Since 1991, he has been a Researcher in Research and Development Laboratory, Korea Telecom (KT), Korea, where he engaged in the areas of Broadband Network and Next Generation Network. His current research interests include aggregate flow control, edge-based QoS provisioning.

Chong Song received his B.S. and M.S. degrees in Control and Instrumentation Engineering from the Seoul National University, Seoul, Korea, in 1988 and 1990, respectively, and his Ph.D. degree in Electrical PLACE and Computer Engineering from the University of PHOTO Texas at Austin in 1995. From 1994 to 1996, he HERE was a Member of Technical Staff in the Performance Analysis Department at AT&T Bell Laboratories, Holmdel, New Jersey, USA. He is currently an Associate Professor with the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea. His research interests are in high-speed communication networks, highperformance switching/routing systems, multimedia networking and performance evaluation. He has published more than 25 papers in international journals and conferences and holds three U.S. patents with several others pending. He is an Editor and an Associate Publication Editor for the Journal of Communications and Networks. He has served as a Technical Program Committee member of IEEE INFOCOM (‘97, ‘99, ‘03), an Organizing/Program Committee member of PV (‘01-‘04) and Technical Program Co-chair of ICBN ‘04.

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

GATE

Bandwidth allocation scheme

1

1 N N

2

1

1 N N 1 1

2

1

1

1

1 1

REPORT

OLT

7

1

User 1

1

Splitter/ Coupler

2

1

1 N N

NN

2

2

2

2

2

User 2

N

User N

guard time

1

1 N N

2

N N

N

NN

N

ONUs Fig. 1.

Typical topology of EPON

1. Receive REPORT from ONUi

OLT

Rx

1

R g

2

2. Grant the timeslot

N

R

R

3. Update SEI

G

(a) In the case of SEI

now + RTT1

1. Receive REPORT from ONUi

R

g G

Tx

Waste! N R g

2 R

2. Grant the timeslot

1

R

g

4. Send GATE to ONUi

now

new SEI

SEI

Rx

ONUi

R

1

Tx

1

new SEI

SEI

Rx

Rx

g

R

4. Send GATE to ONUi

now

OLT

1

G

Tx

ONUi

g

3. Update SEI

G

1

Tx

R

RTT1

(b) In the case of SEI < now + RTT1

i Fig. 2.

Data from ONUi

G

GATE message

R

REPORT message

GATE scheduling: (a) In the case of SEI ≥ now + RT Ti , (b) In the case of SEI < now + RT Ti

g

guard time

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

8

New coming ONU4

1st cycle 2nd cycle

Limited by Gmax

R1=inf

1

g

2, waste

g

R2=0

3, waste

g

R3=0

Limited by Gmax

R4=inf

g

Limited by Gmax

R1=inf

1

4, waste

g

2, waste

g

R2=0

3, waste

g

R3=0

4

R4=inf

Fair

g

(a) Fixed scheme New coming ONU4

Limited by Gmax

1st cycle

R1=inf

1

g

g

R2=0

g

R3=0

Limited by Gmax

2nd cycle

1

g

R4=inf

Limited by Gmax

R1=inf

g

g

R2=0

g

R3=0

4

g

R4=inf

Fair

(b) Limited scheme New coming ONU4

1st cycle 2nd cycle 3rd cycle

F1=4Gmax

1

g

R1=inf

g

R2=0

g

R3=0

R4=inf

g

F4=0, but minimum guarantee Gmax

F1=4Gmax

1

g

R1=inf

g

R2=0

g

R3=0

F1=3Gmax

4

R4=inf

g

F4=Gmax

1

R1=inf

g

g

R2=0

g

R3=0

4

R4=inf

g

Unfair

(c) Proposed scheme 1 New coming ONU4

1st cycle 2nd cycle

s1=0,s2=1,s3=1,s4=1

1

R1=inf

g

R2=0

g

R3=0

g

s1=0,s2=1,s3=1,s4=0

1

g

R3=0

g

R4=inf

g

R4=inf

g

s1=0,s2=1,s3=1,s4=0 R1=inf

g

R2=0

4

(d) Proposed scheme 2 i Fig. 3.

Data from ONUi

R

REPORT message

g

Fairness example: (a) Fixed scheme, (b) Limited scheme, (c) Proposed scheme 1, (d) Proposed scheme 2

guard time

Fair

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

9

1 N=32, Gmax = 125us, guard time = 10us

0.9 0.8

Limited, Fixed for k = 1 Proposed schemes for all k

Utilization

0.7

k=10

0.6

Limited

0.5

k=100

0.4

k=10

0.3 0.2

Fixed

k=100

0.1 0

Fig. 4.

0

0.8

0.7

0.6 0.5 0.4 Offered load

0.3

0.2

0.1

1

0.9

Utilization comparison under uniform and non-uniform traffic conditions

guard time

i

OLT

R

i+1

Waste! R

i-1

i

R

G

ONUi RTTi i Fig. 5.

Data from ONUi

RTT effect: timeslot waste due to round-trip time

G

GATE message

R

REPORT message

IEEE/OSA JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. XX, NO. XX, NOVEMBER 2004

10

c Gmax

OLT

th

guard time

1

R

2

3

R

R

N

R

1

N

R

1

G

ONU1

RTT

c

(a)

c

NGmax

OLT

1

R

2

R

3

th

R

G

ONU1

RTT

c

(b) Fig. 6.

ρth calculation (a) Limited service scheme, (b) Proposed schemes

§L¨0©+ª«ªU¬ ­ ® ¬¯I°



tr s!uv s w vyxyz s w v vIy‚  e{ | |~} rI s €   ! "$#%&' ( B'CEDGFIH

LJ KNMAOAPRQ OISUTWVNX X YGZ+VS[Q Y\M^]Z$T_N`a_I] bced f'gAhLijAd k d lmg+h\inpoAiAn$q Fig. 7.

Simulation model using OPNET

 

)+*-, .0/1(2 3 2 4 5-6 7 6859:.A@'6='* ƒA„†… ‡0ˆ‰(Š ‹ Œ … Ž Žm(‡0‡