Hybrid Resource Reservation Mechanism for Optical Burst Switched

Document Sample
Hybrid Resource Reservation Mechanism for Optical Burst Switched Powered By Docstoc
					            Hybrid Resource Reservation Mechanism for Optical Burst Switched

                                      Huifang Kong and Chris Phillips
                                   Department of Electronic Engineering
                                     Queen Mary, University of London
                          Email: {huifang.kong, chris.phillips}

Abstract           This paper proposes a novel hybrid resource reservation mechanism that can operate with
existing Optical Burst Switching (OBS) architectures, providing an efficient infrastructure for multi-service
“bandwidth on demand” transport. It proposes deploying bursty traffic in a hybrid fashion where implicitly
predicted and explicitly pre-booked traffic are dynamically allocated reserved end-to-end paths, inheriting the
spirit of conventional wavelength routing; whilst, the non-predicted traffic is transmitted via classical OBS
reservation mechanism(s) with the best efforts support. The complete network structure is presented and an
explicit coarse-interval load-balancing prior reservation strategy is also described in detail. Simulation results
reveal the performance of the proposed work by examining the blocking probability, end-to-end delay, and the
wavelength deployment characteristics. The encouraging results should provide stimulation for further work on
optimal traffic placement, QoS provisioning, and various a-priori resource reservation strategies.

Keywords: OBS, Wavelength Routing, Resource Reservation

1. Background

     Wavelength Division Multiplexing (WDM) has emerged as a core backbone transmission
technology with the ability to carry bits at gigabit per second speeds. In current commercial WDM
networks, Optical Circuit Switching, also called Wavelength Routing (WR), is deployed as the
dominant switching technology, where a dedicated end-to-end lightpath is setup for each connection.
Current quasi-static WR places the lightpaths statically, assuming that all the traffic demands are
known in advance and remain constant once in place. However, the rapid growth of IP traffic has not
only changed the volume, but also the nature of the traffic being carried into highly varying bursty
flows. This change has triggered considerable research effort into finding efficient protocols and
switching technology for next generation WDM networks.

      Optical Burst Switching (OBS) is one of the viable candidates for the next generation optical
backbone [1] providing a buffer-less transportation medium. The main feature of OBS is the separated
transmission of the burst data and the corresponding control information (called Burst Head Packet
(BHP)). Each burst is preceded by its own BHP, which travels slightly ahead, configuring the switches
and reserving a wavelength path for the upcoming burst. When the BHP is processed, it undergoes
optical-electro-optical (OEO) conversion at each hop, whilst the data burst is transparently switched in
all-optical manner.

      There is still scope for further exploration in the OBS realm. The most important three issues to
be addressed are delay reduction, burst loss reduction, and QoS provisioning. In classical OBS,
resource reservation is initiated upon the arrival of the burst. This intuitively introduces extra delay, as
the burst has to wait at the ingress point for an offset of sufficient duration to allow the core switches
to be configured before the burst is emitted. Secondly, as the typical burst length in classical OBS is of
the order of tens of kilobytes, one-way reservation is preferred to improve the overhead efficiency.
However, one-way reservation results in un-assured best-effort service, and the most popular OBS
scheme JET using one-way reservation exhibits high burst blocking-rates at high traffic loads. This
problem becomes more acute if the network supports multiple service levels. Many QoS solutions,
such as priority-based OBS [2], can only improve the performance of high-priority traffic by
sacrificing the performance of low-priority traffic and, thus, cannot improve the overall network

      A new approach called Wavelength-Routed OBS (WR-OBS) has been proposed [3]. It deploys a
centralized node to dynamically schedule and place the subscribed bursts by going through an efficient
routing and wavelength allocation algorithm and setting up a small duration lightpath (in the range of
microseconds) for each burst. This approach overcomes the drawback of quasi-static wavelength
routing by shrinking the connection granularity to the burst level, thus providing a dynamic and near-
optimal solution. However, the major concern of this approach is its applicability in large backbone
networks, where the scalability of this centralised solution is questionable under high traffic load

      Nevertheless, inspired by the idea of WR-OBS, this paper proposes a viable enhancement to the
OBS architecture where the burst traffic is transmitted in a hybrid fashion between wavelength routing
and packet (burst) switching [4][5]. The core principle of proposed scheme is that, network-planning
components proactively reserve wavelength resources for implicitly predicted or explicitly pre-booked
future traffic. This prior reservation concept inherits aspects of wavelength routing, providing end-to-
end guaranteed paths 1 ; and an efficient routing and wavelength allocation algorithm, to place the
anticipated traffic flows appropriately across the network resources. The actual traffic flows will be
principally delivered via these reserved paths; however, if the demand is exceeds the reserved path
capabilities, the excess traffic can be delivered via classical OBS signalling with its inherent risks.
This architecture takes advantage of traffic prediction to improve the wavelength routing efficiency.

      In the remaining part of this paper, the basic network structure is first presented by describing the
functionality of key network components; it then describes the implementation of a coarse-interval
load-balancing reservation mechanism. Its characteristics in terms of end-to-end delay and blocking
rates under different prediction accuracies and traffic loads are examined; and, finally, some
conclusions are drawn.

2. Proposed Network Structure

                                                      Central Network Planning Node

                                   SP1                                                              SP3


                                              CNO Edge Node                 CNO Core Node

                                              Carrier Network Operator (CNO)

                                          Figure 1 –Proposed Network Architecture

      The proposed network structure is shown in Figure 1. The WDM backbone is operated by a
Carrier Network Operator (CNO), with various Service Providers (SPs) placed at the edge. For
scalability purposes, all the complexity is mainly placed at the boundary between the CNO and the
SPs, allowing the core optical switches to be relatively simple devices.

    The term path is loosely applied, as it does not simply imply the existence of a spatially and temporally invariant end-to-
    end connection. The actual path may effectively only exist across part of end-to-end path at a given point in time, as it is
    established and then released as the burst moves across the core network.

2.1 Service Provider

      Service Providers at the edge of the carrier network manage customer traffic, and are responsible
for burst assembly, burst delivery, and QoS selection. The main challenge for them is to adopt an
appropriate strategy of resource subscription from the CNO and allocation of these resources across
their customers.

      In the proposed approach, each SP (as shown in Figure 2) consists of two key parts –a resource
reservation mechanism and a resource allocation unit. The prior resource reservation centre acts as a
resource request agent, liaising with the CNO. It accepts explicit future pre-booking from customers
via a long-term pre-booking clerk ; i also predicts future end-to-end traffic demands based on
historical traffic patterns and other implicit means available to it. The pre-booking and implicit
prediction information will then used to formulate reservation requests stipulating parameters such as
reservation time, duration and pre-emption level. These requests are then sent to the central network-
planning node on the CNO side to ask for suitable resources to be reserved, as required2 . The resource
allocation part of SP is responsible for feeding the actual burst traffic into the reserved resources at the
pre-booked time(s) they become available. In the case when the reserved resource is insufficient, the
resource allocation part will typically arrange to issue the classical OBS signalling to de liver those
bursts on-demand. In terms of QoS provisioning, it is proposed to set aside more resources for loss
sensitive traffic, such that the higher class of traffic has more reserved resources thus lowering the risk
of burst blocking / loss. Since QoS provisioning is not the major concern of this paper, it will not be
described in detail.

                                 Long-term Pre-booking Clerk

                                                                                                      Central Network Planning
                                                                                                      Node owned by CNO
        Burst Assembly Buffer       Implicit Traffic Predictor    Prior
           Destination 1                                                          Prior Resource Reservation Request

                                                                   Prior Resource Reservation Confirmation


           Destination (N –1)
                                               Resource                                                Optical Fabric owned by
                                               Allocation                                              CNO Edge Node


                                        Figure 2 –Structure of Service Provider

      The motivation the prior resource reservation scheme proposed here is to be able to help service
providers to dynamically reserve network resources based on explicit customer pre-booking and
estimated demand. However, determining the frequency of prior reservations, including their start time
and duration, remains a challenging task especially for implicit predicted traffic, where the source-
destination traffic matrix is non-stationary. For example, a service provider can predict with varying
degrees of certainty the average traffic load in next one hour, or in next five minutes; the service
provider can also alternatively predict the detail of each burst within the next few seconds. Together
with the uncertainties brought by the traffic arrival processes themselves, it will be interesting to
compare the benefits of different reservation strategies, though this topic remains the focus of ongoing
    Note, factors such as the path are not accessible to the SP. It is up to the CNO, how or whether the resources will be
    deployed to meet the reservation request.

2.2 Carrier Network Operator (CNO)

      In the CNO, the major concern is how to place the traffic in an optimal way such as to maximise
the traffic volume carried, and avoid situations there certain parts of the network are unnecessarily
congested while other parts are under-utilised.

       The CNO accepts and handles the prior resource reservation requests via a central network-
planning node. The central network planning node (as shown in Figure 3) collects all the prior
reservation requests from SPs into N*(N-1)*C reservation request queues, where N refers to the
number of edge nodes, and C refers to the number of priority service classes. It then tries to optimally
place the subscription requests by running an efficient routing and wavelength allocation algorithm, or
formulating the provisioning as an Integer Linear Programming (ILP) problem. Based on the results, a
sequence of acknowledgement-required two-way reservation will then be issued to finally confine the
reservation. Given that these requests pertain to future requirements, the optimisation algorithm does
not need to operate “               .
                         on-the-fly” Indeed, depending upon the remoteness of the reservation times,
iterative of differing placement mechanisms could be supported.

               Prior Resource
               Reservation Request

                 Destination 1



                                                    Network           Reservation
                    CoSc                                              Request
                 Destination N*(N – 1)                               Acknowledgement


                        Figure 3 –Structure of Central Network Planning Node

      Apart from the central node, the CNO infrastructure also has the ability to support classical OBS
reservations. As classical OBS requires topological knowledge at the ingress, each edge node
maintains a periodically updated link state database and a source-routed forwarding table, which
specifies routes from a source to each destination egress point. The source-routed forwarding table is
updated in response to changes in the link state database. The link state database is also influenced by
knowledge of confirmed reservations from the central network-planning node.

3. Implementation of Coarse Interval Average Load Prior Reservation Strategy

3.1 The Coarse Interval Average Load Prior Reservation Strategy

     As it has been mentioned that service providers can adopt various prior reservation strategies in
terms of subscription granularity and composition, this paper focuses on the implementation of a
coarse-interval load-balancing prior reservation strategy.

      In this strategy, a service provider predicts the average end-to-end traffic load over a large time
interval, such as every hour, and puts the minimum wavelength requirements into the prior reservation
request. The prior resource reservation then has to reserve required amount of lightpaths for the one-
hour duration. The minimum end-to-end (EtE) wavelength requirement is calculated as the following
formula, where ceil means the smallest integer that is greater than or equal to the real value in the

                                                  predicted EtE load (Gb / s) 
                                                  wavelength rate (Gb / s) 
                    wavelength requirment = ceil                              
                                                                              

      The reason for developing a coarse-interval load averaging prior reservation scheme is that, with
the current state of prediction technology [6], it is much easier to forecast large interval average traffic
load due to the relatively stable daily traffic patterns; whilst it is very difficult to predict the
characteristics of each burst.

     On the CNO side, the traffic placement is based on a modified form of Dijkstra’ algorithm,
where the link weight is increased once the reservation is placed on the link. This facilitates load
balancing, but it does necessarily yield an optimal solution.

3.2 Wavelength Assignment without Wavelength Conversion

      Another important issue needs to be noted is that in order to be more realistic, the wavelength
continuity constraint is applied in the current implementation. Therefore, all the prior reservations
correspond to continuous single -wavelength lightpaths. This raises the wavelength selection problem
once the path is determined. In the current implementation, for traffic that can be carried on prior
reserved resources, it employs the latest available unused channel with void filling (LAUC-VF) [7]
algorithm to select the wavelength on the first link. Because a prior reservation is end-to-end
guaranteed, the resource availability along the whole path can be guaranteed if the resource is
available on the first link. Conversely, in classical OBS reservations with the wavelength continuity
constraint, the optimal wavelength selection on the first link can hardly bring significant benefits,
because it is a one-way best effort reservation and the selected optimal wavelength on the first link can
be occupied by other bursts along some later links. Therefore, the current implementation chooses to
randomly select a wavelength at the ingress node for classical OBS reservations.

4. Results and Discussion

                                                                                     800      NY
                     WA                                  2400
                                                                                 800           300
             1100                                                                     500
                       1600                                                             500        NJ
                                                                         700    PA
                                                      MN                                             300
                              UT         CO     800           700
            CA1                    600
                       1000                                                    900            MD
                    CA2                                        TX                    GA
                                         2000                            1200

                                Figure 4 –Simulation Topology (in km)

      Simulation results of the proposed approach, using the coarse interval load-balancing prior
reservation strategy, have been collected via OPNET™ simulation framework. They present the
performance of the scheme in terms of blocking probability, end-to-end delay, and wavelength
deployment, under different prediction-accuracy levels and various end-to-end loads. It employs a
time-based burst consolidation mechanism in the service provider’ burst assembly section, where
bursts are emitted at a constant time interval. The burst for each source-destination pair are modelled
using on-off model with negative exponential random on and off periods. The average duration of the
on period is 0.96 seconds, and that of the off period is 1.69 seconds. The burst generation rate within
the on period is 0.05 seconds for each burst (equalling 50 milliseconds per burst). The tests are carried
out using a 14-node NSFNET topology as shown in Figure 4. The transmission rate of each

wavelength is 10Gbps, and there are 64 wavelengths on each link. Across the NSFNET network, the
burst traffic is generated at a specified load for all the source-destination pairs, where the traffic matrix
is fixed and balanced.

4.1 Blocking Probability

      Figure 5 shows the blocking probability versus prediction accuracy level with various end-to-end
loads, where the accuracy level is defined as the ratio of predicted end-to-end (EtE) load to the actual
end-to-end (EtE) load.

                                                                        predicted EtE load
                                                    accuracy level =
                                                                         Actual EtE load

      Thus, the amount of prior reservation increases with increasing accuracy levels given a fixed
actual offered load. The minimum wavelength reserved for each source destination pair through prior
reservation equals:

                                                         accuracy level * actual EtE load ( Gbps ) 
                 EtE prior reserved wavelength s = ceil 
                                                                                                   
                                                                wavelength rate ( Gbps )           


    Blocking Probability






                                0 1 2 3 4 5
                                                                      Accuracy Level
                                         Load=5Gb/s Load=10Gb/s Load=15Gb/s Load=20Gb/s Load=25Gb/s
                                         Load=30Gb/s Load=35Gb/s Load=40Gb/s Load=45Gb/s Load=50Gb/s

                                                        Figure 5 –Blocking Probability

      The results show that the blocking probability is not only affected by the prediction accuracy, but
also by the offered load. When each source destination offered load is as low as 5 Gbps, the blocking
probability decreases with the increase of accuracy level. This is because the offered load is very low;
the prior resource reservation does not cause contention, although the prior subscription requirements
are far beyond the actual m    inimum wavelength requirements. In this case, as more resources are
reserved via prior reservation, more bursts can be delivered via prior reserved end-to-end guaranteed
resources, thus resulting in lower loss. However, with the increase in offered load, contention occurs
within the prior reservations, as well as within the classical OBS reservations. That is why the
blocking probability curve exhibits a minimum, with high end-to-end offered loads. It also indicates
that with a high offered load, it is very important to maintain the prediction accuracy at a reasonable

level (at about 1.0 to 1.4 in this experiment set), so as to avoid unnecessary prior reservations, and
consequentially reduce the losses.

     On the other hand, the blocking probability of pure cla ssical OBS is shown by the first point of
each curve, as with zero prediction accuracy level, no prior reservation takes place in the core network.
Therefore, this result also shows that by proper control of the prediction accuracy level, the proposed
hybrid reservation architecture is able to obtain lower blocking probability than the pure classical OBS.

4.2 End-to-End Delay

          The end-to-end delay in classical OBS can be formulated as follows:
                                         Dclassical = Daggregation + offsettime + Dtransmission + D propagatio

     where the end-to-end delay of classical OBS Dclassical is mainly caused by Daggregation , the burst
aggregation delay (0.05/2 = 0.025seconds in our simulation); offset time (100 microseconds for each
hop BHP reservation operation in simulation); Dtransmissi , the transmission delay (determined by the

burst length and wavelength rate); and D propagation , the propagation delay (determined by path, fibre
length, and fibre propagation speed).

    End-to-End Delay (s)







                                 0 1 2 3 4 5
                                                                       Accuracy Level
                                      Load=5Gb/s Load=10Gb/sLoad=15Gb/sLoad=20Gb/sLoad=25Gb/sLoad=30Gb/s

                                                          Figure 6 –End-to-End Delay

     In this proposed scheme, as the prior lightpaths are setup before the bursts arrive, the end-to-end
delay for those bursts using prior reserved lightpaths can be formulised as:
                                             D prior reservation = Daggregation + Dtransmission + D propagatio

      As there is no need to delay the burst for an offset time to allow the resource reservation by BHP,
the end-to-end delay can actually omit the offset time. However, in our results (as shown in Figure 6),
deduction of the end-to-end delay has very little effect of prediction accuracy, whilst it increases with
the increasing end-to-end offered load. This is because the offset time is of the order of microseconds,
whilst the propagation time and aggregation time are both in the range of milliseconds; the effects of

offset time reduction are marginal. Moreover, because the increase of offered load typically results in
larger bursts3 , the end-to-end delay is therefore increased by the longer transmission delay.

4.3 Wavelength Deployment Situation

     It is also important to examine the wavelength deployment characteristics over the whole
network, which can be demonstrated by the probability distribution function of the average number of
wavelength deployed on each link, as shown in Figure 7 and Figure 8.
       In Figure 7, it shows the wavelength deployment situation at a fixed prediction accuracy level of
0.8, and with different levels of end-to-end offered load. The different colour used for each blocking
probability point (in the line graph) also relates to the similarly coloured wavelength deployment
distribution that arises at the same offered load. It shows that with the increase of offered load, the
probability distribution curves move to the higher end of the wavelength deployment, however, the
curves start to spread into a wider range of deployed wavelengths. This reveals imbalances in the
traffic placement; the modified Dijkstra algorithm used in this implementation is most probably not
optimal, causing unnecessary blocking at high loads.

                                                                  Load (Gb/s) of each Source-Destination Pair
                                                                         (Wavelength Rate = 10Gb/s)
                                          5                       15                   25                  35                     45
                               2.50%                                                                                                   60%
                                              Probability Distribution Function of Average      Blocking Probability versus Load of
                                              Number of Deployed Wavelengths on each            each Source-Destination Pair
                                              Link                                                                                     50%

                                                                                                                                            Blocking Probability
          Probability of WiU




                               0.00%                                                                                                   0%



                                                          Average Number of Wavelength Deployment per Link (WiU)

                                        Load=5Gb/s Load=15Gb/s Load=25Gb/s Load=35Gb/s Load=45Gb/s Blocking Probability

                               Figure 7 –Wavelength Deployment with a Fixed Prediction Accuracy Level at 0.8

      Figure 8 shows the wavelength deployment behaviour obtained by fixing each source destination
offered load at 25 Gbps. Similar as in Figure 7, Figure8 also uses different colours to link the blocking
probability and the corresponding wavelength deployment distribution at the same accuracy level. It
can be clearly seen that as the accuracy level increases, the corresponding probability distribution
curve shows that increasing numbers of wavelengths are deployed. However, as the level of undue
pre-booking rises beyond a figure of 2.0 in this case, less wavelengths are actually deployed; the
remainder are unnecessarily reserved. On the other hand, the degree of spread of each curve remains
unchanged. This means that there is still scope for improvement of the core resource allocation

    In the ON-OFF burst generation model, the burst generation rate parameters are fixed. Therefore, under high traffic load,
    more packets are pushed into one burst, which results in a larger burst length.

                                                                                Accuracy Level
                                       0                  1                 2                    3                  4                  5
                            1.40%   Probability Distribution Function of Average                     Blocking Probability versus Load of
                                    Number of Deployed Wavelengths on each                           each Source-Destination Pair
                                    Link                                                                                                         50%


                                                                                                                                                      Blocking Probability
       Probability of WiU




                            0.00%                                                                                                                0%









                                                   Average Number of Wavelength Deployment per Link (WiU)
                                        Accuracy Level=0.0   Accuracy Level=1.0     Accuracy Level=2.0                      Accuracy Level=3.0
                                        Accuracy Level=4.0   Accuracy Level=5.0     Blocking Probability

Figure 8 –Wavelength Deployment with Fixed End-to-End Offered Load at 25 Gbps for each Source-
                                      Destination Pair

5. Conclusion

      This paper proposes a novel prior reservation OBS mechanism that can operate over existing
core OBS architectures, providing a self-optimising infrastructure for multi-service “ bandwidth on
demand”transport. The results show the potential and limitations of currently proposed coarse interval
load-balancing prior reservation strategy. The remaining challenge and future work will be to devise
an efficient traffic placement mechanism, and to explore the performance of QoS provisioning under
the proposed scheme. Furthermore, it will be interesting to explore other pre-booking strategies,
possibly tailored to reservations across particular timescales.


[1]   Verma, S.; Chaskar, H.; Ravikanth, R.; “Optical burst switching: a viable solution for terabit IP
      backbone”, Network, IEEE , Volume: 14 , Issue: 6 , Nov.-Dec. 2000, Pages:48 –53
[2]   H.L. Vu, and M. Zukerman; “  Blocking probability for Priority Classes in Optical Burst
      Switching Networks”, IEEE Communications Letters, vol. 6, issue 5, pg. 214-216, May 2002.
[3]   Duser, M.; Bayvel, P.; “Analysis of a dynamically wavelength-routed optical burst switched
      network architecture”, Journal of Lightwave Technology, Volume: 20, Issue: 4 , April 2002
      Pages:574 –585
[4]   Huifang Kong and Chris Phillips, “Intelligent Pre-Booking for Optical Burst Switching”  ,
      INOC'2003 - International Network Optimization Conference, Paris (Evry), France, October
[5]                                    Pre-booking for Optical Burst Switching”, Proceedings of the
      Huifang Kong and Chris Phillips, “
      London Communications Symposium 2003 (LCS), London, UK, September 2003.
[6]                                     A
      Groschwitz, N.K.; Polyzos, G.C.; “ time series model of long-term NSFNET backbone traffic”,
      Communications, 1994. ICC 94, SUPERCOMM/ICC '94, Conference Record, Serving
      Humanity Through Communications. IEEE International Conference on , 1-5 May 1994
      Pages:1400 - 1404 vol.3

[7]   Tan, M. Gurusamy and B. Wang; “  Burst Rescheduling with Wavelength and Last-hop FDL
      Reassignment in WDM Optical Burst Switching Networks”, In Proceedings of IEEE ICC,
      volume 2, pages 1448-1452, 2003.


Shared By: