You are on page 1of 5

A DYNAMIC BANDWIDTH ALLOCATION SCHEME WITH TWO THRESHOLD

REPORTING AND STRICT PRIORITY SCHEDULING FOR MILTI-SERVICE EPONsI

Lei Bin, Xue Chen, Rong Gao, Xiaoguang Lu

Beijing Univ. of Posts & Telecom


ABSTRACT upstream direction, all ONUs share the same uplink
transmission channel. Therefore, there must be some
Ethernet-based Passive Optical Network (EPON) is mechanism to prevent data transmitted by different ONUs
considered as one of the most promising solutions for the colliding with one another in upstream direction.
next generation broadband access network. Dynamic Assigning each ONU non-overlap time slot is an available
bandwidth allocation (DBA) is a key factor to fully solution. In this way, each ONU can transmit data in its
exploit the potential of EPONs. Furthermore, DBA is assigned time slot while other ONUs keep silent. While
needed to support QoS for differentiated service assigning each ONU fixed time slot is simple, it cannot
efficiently. A simple DBA scheme called TTSP is adapt to bursty traffic and will lead to degradation of the
proposed in this paper. TTSP specifies two thresholds to system performance. Thus, a dynamic bandwidth
eliminate unused time slots completely to improve allocation (DBA) scheme, which allocates bandwidth to
throughput of EPON systems and maintains a counter for ONU in accordance to its demand, is desirable.
each ONU to guarantee fairness of bandwidth allocation
among ONUs. Simulation results show that TTSP can 2. RELATED WORK
achieve high throughput, guarantee the rigid delay bound
of high priority traffic which is delay-sensitive and ensure IEEE 802.3ah Task Force has developed Multi-Point
the fairness of bandwidth allocation. Moreover, TTSP is Control Protocol (MPCP) [1] to manage the control
simple and easy to be implemented. It has already been messages that are exchanged between OLT and ONU.
implemented by FPGA in our multi-service EPON However, MPCP just provides a framework for the
systems and test results are consistent with simulation communication between OLT and ONU, it does not
ones. specify any particular bandwidth allocation scheme.
Reference [2] proposed an Interleaved Polling with
1. INTRODUCTION Adaptive Cycle Time (IPACT) scheme based on MPCP.
The limited service type of IPACT DBA in Reference [2]
Today, subscribers require more and more multimedia exhibits good performance by predefining a maximum
applications such as voice, data and video, which transmission window size. In limited service DBA, OLT
consume a lot of bandwidth. However, the access network grants ONU its request number of bytes but no more than
has been a bottleneck to provide enough bandwidth to the predefined maximum transmission window size.
support various kinds of applications. What is more, However, OLT does not distinguish the traffic from
subscribers hope to enjoy different kinds of applications different services, nor does it concern with how an ONU
with different quality of service (QoS), which is beyond uses its assigned time slot to transmit traffic of different
the ability of traditional access network technologies. services. In addition, due to the fact that Ethernet frames
Therefore, various solutions have been put forward to cannot be fragmented, when a frame does not fit into the
meet the increasing bandwidth demand and support time slot, it will have to wait for the next transmission
differentiated services. Among these solutions, EPON is opportunity. In this case, a certain amount of assigned
considered as one of the most promising solutions to time slot remains unused, which in turn leads to
achieve both goals with relatively low cost. degradation of throughput. References [3] [4] [5] have put
An EPON is a point to multipoint network which forward several DBA schemes to satisfy different QoS of
consists of one optical line terminal (OLT) and multiple multiple services and raise throughput as well. Reference
optical network units(ONU). In the downstream direction, [3] proposed a two-stage queuing scheme to eliminate
OLT broadcasts Ethernet frames to all ONUs, and ONU light-load penalty and unused time slots. But this scheme
selectively deliver some to its end users, while in the will result in the increase of packet delay of high priority

1 This paper is sponsored by National 863 High Technology Projects of China(Project No:2001AA122021 and 2004AA122320)

1-4244-0463-0/06/$20.00 ©2006 IEEE


service. Reference [4] proposed FQSE scheme that can .P I Queue Ethemet
f

Ethernet
decrease unused time slots. However, it cannot eliminate Frame 4 1 Frame #:
-1
600 bytes byles
unused time slots completely. D-CRED proposed in 5M

Reference [5] can eliminate unused time slots completely


and improve throughput significantly. However, D-CRED
|tliteriiet
1P2 Queue17tIne #3 Ethemel rtl[
is relatively complex due to its dynamic thresholds. 400 bytes 80 bytes
In order to achieve high throughput, appropriate QoS Logic Eh
for different services and low complexity, a Two co-mbi'tletion Frallne 4:1
e
FrameT] 42 |F
Threshold reporting and Strict Priority scheduling DBA of PI a"iid P2
scheme is proposed here. Instead of dynamic thresholds,
two static thresholds are employed in the proposed
algorithm. In the following sections, the Two Threshold
reporting and Strict Priority scheduling DBA algorithm,
which is called TTSP for short, is investigated in detail.
Figure 1 An example of calculating SI and S2
3. THE MECHANISM OF TTSP SCHEME
3.2. Inter-ONU bandwidth allocation at OLT side
The mechanism of TTSP DBA scheme includes ONU
reporting, OLT granting and ONU scheduling processes. TTSP is based on the limited service type of IPACT DBA
In order to guarantee TDM service's QoS requirement, for inter-ONU bandwidth allocation at OLT side.
TDM service is handled in a specific way in the TTSP After receiving REPORT message in which S2 is
scheme. larger than SI, OLT will check the corresponding counter
value of the ONU. If it is negative, OLT will issue SI to
3.1. ONU reporting the ONU for grant, and if it is non-negative, OLT will
issue S2 for grant. When SI is equal to S2, OLT will not
For the sake of eliminating unused time slots completely, check the counter value and will grant the transmission
two thresholds denoted by TI and T2 in unit of byte are window size SI to the ONU directly.
specified in each ONU, and the relationship between the The initial counter value of an ONU is assumed to be
two thresholds is as following: zero and it changes only if S2 is larger than S1. In the
following, an assumption that S2 is always larger than SI
T2=Ti+M (1) is made. Suppose that in the first cycle, the counter value
Where M is a constant positive integer. The value of TI in
a specific ONU is determined by the contracted assured of ONU A is zero, thus, OLT will issue S2 for grant to
bandwidth of this ONU and the reference cycle time. ONU A. Then the counter value of this ONU will
Reference cycle time is set in advance and it equals the decrease into (TI-S2) which is negative. This negative
sum of all ONUs' transmission time when all the granted counter value implies that the actual transmission window
transmission window sizes are TI. Moreover, Maximum size that OLT issues for grant is larger than TI which is
transmission cycle time equals the sum of all ONUs' the amount that ONU A should get according to its
transmission time when all the granted transmission contracted assured bandwidth. Therefore, in the following
window sizes are T2. The value of M should be carefully few cycles, OLT will issue SI instead of S2 for grant to
chosen so that it is large enough to accommodate a frame ONU A to offset the exceeding part of transmission
and will not make maximum transmission cycle time too window size issued in the first cycle and the negative
large. counter value of ONU A will increase by the amount of
An ONU reports its buffered Ethernet frame (TI-SI) in each cycle. OLT will not issue S2 for grant to
boundaries denoted by S I and S2 that are just below these ONU A until the counter value of ONU A turn non-
two thresholds separately. An example is shown to negative again. Non-negative counter value of ONU A
illustrate the process to calculate SI and S2 in figure 1. In implies that the sum of transmission window sizes that
this example, there are two traffic priority queues namely OLT issued in previous cycles to ONU A is less than the
P1 and P2 in an ONU with P1 being the higher level amount ONU A should get, so when the counter value of
traffic and P2 being the lower one. S2 can be equal to SI ONU A turn non-negative, OLT will issue S2 for grant to
when the sum of sizes of all the frames that buffered in ONU A to offset the deficit part and the counter value will
the ONU is equal to and less than T1. In addition, SI and decrease by the amount of (S2-TI). In this way, the
S2 contained in the REPORT message of an ONU are counter value of ONU A fluctuates around zero. By
changed each cycle. granting based on the fluctuating counter value, TTSP
ensures that each ONU gets its fair bandwidth share. The
c code description of OLT granting when it receives
REPORT message from an ONU is given as the following:
if(S 1=S2)
Grant Size=S 1;
if(S 1<S2)
{
if(counter_value<=0)
Grant Size=S2;
else
Grant Size=S 1;
counter_value=counter_value+T 1 -Grant-Size;
} Figure 2 Average packet delay Vs network offered load
3.3. Intra-ONU Scheduling at ONU side

In order to support multiple services and guarantee their


different QoS requirements, different service classes have
to be defined and different service data are buffered in
separate queues.
Here services are classified into three priority levels
as defined in [2].
The high priority service P0 supports a traffic class
for so-called TDM service applications that is of constant
bit rate and require bounded end-to-end delay and jitter
performance. It is used for voice service, El lease line
service and so on. Figure 3 Maximum delay of PO traffic Vs network offered
The medium priority service P1 supports a traffic load
class for applications that are not delay sensitive but
require guaranteed bandwidth. It is used by non-real time
services of variable bit rate.
The low priority service P2 supports applications that
require neither delay bound nor bandwidth guarantees. It
is a best-effort service class.
When an ONU receives its GRANT message sent by
OLT, it will schedule the Ethernet frames in its multiple
xR
3 9.5

queues in strict priority. In strict priority scheduling, ONU


Fiue4Trogpto
loa ..d
TTS an
,'
NTS Vs newr fee
serves the higher priority queues to exhaustion before
serving a lower priority queue. In particular, in order to
eliminate the light-load penalty presented in [3], only the
reported frames are eligible to access the link.
3.4. P0 class traffic handling

In ONU-Level scheduling, only those frames that are


reported are eligible to access the link. Thus, another
subtle problem arises. If a frame arrives immediately after
the generation of REPORT message, it will have to wait
for two cycles for its transmission opportunity. This
problem will result in the increase of delay which is not
acceptable for high priority traffic P0 which require rigid
delay bound.
In order to reduce delay of P0 class traffic, OLT
allocates fixed time slot for P0 traffic in each ONU every
cycle. And this fixed time slot is large enough to Figure 5 Fairness factor Vs network offered load
accommodate all the packets in P0 queue or even larger.
Since P0 class traffic is of constant bit rate, the length of
the fixed time slot can be calculated according to the
maximum transmission cycle and its bit rate. In this way,
P0 traffic does not participate in the REPORT/GRANT
process defined in MPCP.

4. SIMULATION RESULTS
(I~~
A detailed simulation is conducted to study the
performance of TTSP algorithm. Simulation model is as Figure 6 Set-up for TTSP Performance Testing
following:
-Total GNU number is 64.
-Each GNU has three priority queues (High P0, Medium
P1I, Low P2). P0 traffic emulates ElI connection with
constant bit rate of 2Mbps. P1 traffic consists of VBR
video streams that exhibit properties of self-similar and
long-range dependence. P2 traffic is best-effort traffic
which exhibits the same property as P1.
-Total bandwidth is 1Gbps.
-Reference cycle time TR is 1 ins.
-Contracted assured bandwidth of each GNU is 13.5Mbps.
-M (T2 minus T I) is 800 bytes.
Here M is chosen as 800bytes for two reasons. First
of all, it is very likely that the frame size is below
800bytes [7]. Thus, in most cases, M is large enough to Figure 7 Testing throughput and simulation throughput
accommodate at least one frame. Secondly, in extreme
case where all the ONUs' transmission window sizes are
T2, the maximum transmission cycle time TM is about
1.41ms, which is below 1i5ms specified in ITU-T
recommendation [8]. Therefore the delay bound of P0
traffic can be guaranteed. Moreover, the value of M
should be carefully chosen to avoid head-of-line blocking.
Simulation results of TTSP are shown from Figure 2
to Figure 5. Figure 2 indicates that P0 traffic has the
smallest average packet delay while P2 has the largest
average packet delay. In particular, the average packet
delay of P0 traffic is well below 1 .5ms, indicating that
TTSP can guarantee the QoS requirement of delay-
sensitive traffic P0. Figure 3 further illustrates that even
under the heavy network load, the maximum delay of P0 Figure 8 Testing and simulation average packet delay
traffic is 1.lmis which is below the delay bound 1 Sins.
Figure 4 indicates that TTSP performs much better in
throughput than NTSP, which is based on the combination (2)
of the limited service type of IPACT DBA in OLT level FF'='ji =12.--
granting and strict priority in GNU level scheduling. But 2
NTSP does not have any thresholds. Figure 5 shows that
the fairness of TTSP is very good. Here fairness among Bandwidthi
Allocated (3)
different GNUs is measured by fairness factor [9], Contract Bandwidthi
denoted in FF which is defined as the following:
Where Allocated Bandwidth1 is the average value
of bandwidth allocated to GNU i when GNU i requests
more than its grant, Contract Bandwidth1 is the
contracted assured bandwidth of GNU i, and N is the
number of ONU. The more FF approaches zero, the better We would like to express our deepest gratitude to Dong
the fairness. Liu and Pengfei Gai from GW Technologies CO. for their
valuable advices and suggestions on this work. We are
5. THE IMPLEMENTATION OF TTSP also very grateful to many other people who offered help
to us during our work on TTSP algorithm, although we
Due to its simplicity and effective supporting for multiple cannot list them all here.
services, TTSP has already been put into implementation
by FPGA in our Multi-service EPON systems. In addition, 8. REFERENCES
the performance of the implemented algorithm is tested as
well. [1] IEEE 802.3ah, Carrier Sense Multiple Access with
The experimental set-up is as shown in Figure 6. Collision Detection (CSMA/CD) Access Method and
Though only two ONUs are available, the testing set-up Physical Layer Specifications Amendment: Media Access
can be configured to 64 ONUs on line. In addition, there Control Parameters, Physical Layers, and Management
are two queues corresponding to two service priorities P0 Parameters for Subscriber Access Networks, 2004.
and P1 in each ONU equipment in our test. In order to
compare the testing results with the simulation ones, a [2] G. Kramer, "Interleaved Polling with Adaptive Cycle
new simulation in which each ONU has two queues is Time(IPACT): A Dynamic Bandwidth Distribution
conducted, and the simulation results of throughput and Scheme in an Optical Access Network," Photonic Net.
average packet delay are depicted as dotted line in figure Commun, vol. 4, no. 1, pp. 89-107, January 2002.
7 and figure 8 respectively. It is also worth mentioning
that due to the fact that our testing equipment can only [3] G. Kramer, B. Mukherjee, S. Dixit, Y. Ye, and R.
generate CBR source data, P0 and P1 services are of Hirth, "On supporting differentiated classes of service in
constant bit rate in both test and simulation. EPON-based access network," J. Opt. Networks, pp. 280-
Figure 7 shows that the throughput of testing result is 298, 2002.
consistent with simulation very well. The slight decrease
of tested throughput is due to the ignoring of OAM [4] G. Kramer, A. Banerjee, N. K. Singhal, B. Mukherjee,
overhead in simulation. Figure 8 shows that the tested S. Dixit, and Y.Ye, "Fair queueing with service envelopes
average packet delay is slightly above the average packet (FQSE): A cousin-fair hierarchical scheduler for
delay in simulation. The reason for the difference is that subscriber access networks," IEEE J. Sel. Areas Commun.,
there exists some handling delay in real equipments. vol. 22, no. 8, pp. 1497- 1513, October 2004.
Therefore, in general, testing results are consistent very
well with simulation results. This indicates that TTSP can [5] T.Miyoshi,H Inoue, K.Yamashita, "QoS aware
achieve high throughput, guarantee the delay bound of Dynamic Bandwidth Allocation Scheme in Gigabit-
high priority traffic. Ethernet Passive Optical Networks," vol.1, pp.90 - 94
20-24 June 2004
6. CONCLUSION
[6] A. Adas, "Traffic models in broadband networks,"
To efficiently and fairly allocate bandwidth among IEEE Communications Magazine, vol. 35, no. 7, pp. 82-
different ONUs and guarantee multi-service's QoS, a 89, July 1997.
simple DBA algorithm with two threshold reporting and
strict priority, TTSP algorithm, is proposed. Then, [7] K. Claffy, G. Miller and K. Thompson, "The Nature of
simulation is conducted to validate its effectiveness. The the Beast: Recent Traffic Measurements from an Internet
simulation results show that TTSP can achieve high Backbone," Proceedings of the Internet Global Summit
throughput, provide rigid delay bound for high priority Conference, San Diego, CA, February 2003.
traffic and guarantee the fair allocation of bandwidth
among different ONUs. Furthermore, since TTSP is of [8] ITU-T G.982, Optical Access Network to Support
low complexity and very easy to be put into Services up to the ISDN Primary Rate or Equivalent Bit
implementation, it has already been implemented in our Rates, 1996.
multi-service EPON systems by FPGA. Testing results for
real EPON system with TTSP are consistent with [9] Xue Chen, Meihong Yu, Yang Zhang, Yu Deng, "A
simulation very well. Novel Upstream Dynamic Bandwidth Assignment
Scheme for Ethernet PONs," Proceedings of ICCT 2003,
7. ACKNOWLEDGEMENT Beijing, China, April 2003.

You might also like