Analysis of Impact of Control Plane on High Speed Switching Paradigm

Optical burst switching (OBS) network is viewed as two parallel networks, namely a data and a control network (plane). In this paper, the main focus has been given only on the impact of control-plane on performance of OBS. In high speed networks, several factors such as contention, insufficient offset-time, average burst length and scheduling operations may cause excessive delay in processing of control packets in an electronic core-node controller, thus as a result incur higher blocking probability, inefficient bandwidth utilization etc. In this paper, to address this, an analytical modeling, detailed analysis of control plane and its impact on OBS performance has been provide d . Based on above, optimal performance oriented parameters can be selected such as average burst length, sufficient offset-time duration and scheduling and reservation operations in order to show the impact of control-plane negligible (as compared to data-plane) on OBS based system performance.


OPTICAL BURST SWITCHING
To circumvent the potential bottlenecks of electronic processing in optical packet-type wavelength division multiplexed (WDM) networks, the basic data block to be transferred is macro-packet, called burst, which is a collection of data packets having the same network egress address and some common attributes, like Quality of Service (QoS) requirements.The functional block diagram of an optical burst-switched (OBS) network is shown in Fig. 1, which consists of optical core nodes (routers) and electronic edge nodes (routers) connected by WDM links.Packets are assembled into bursts at network ingress, which are then routed through the OBS network and disassembled back into packets at network egress to be forwarded to their next hops (e.g., conventional IP routers).In a simple OBS network configuration, data-bursts (DBs) and their control packets are transmitted on the same link but on different wavelengths and are separated in time by an offset.It is assumed that each link contains a single control channel, the core node is buffer-less and data bursts are IP-based.Except for the separate transmission of headers and payloads and being switched in different domains, there is no fundamental difference between packet switching and the OBS.However, in the OBS, a burst header must explicitly reserve the switching resources in advance at each hop along the path for its burst payload, while in store-and-forward packet switching; the reservation of switching resources is made implicitly, i.e. when a packet is sent out from an electronic buffer.
The function of the switch control unit (SCU) is similar to a conventional electronic router.The routing processor runs routing and other control protocols for the whole OBS network.It creates and maintains a routing table and computes the forwarding table for the SCU.In arranging the transfer of a data burst and its corresponding BHP in the optical switching matrix and SCU, respectively, the SCU tries to resynchronize the data burst and the burst header packet (BHP) by keeping the offset time.If there are free data and control channels available from these groups, either when the data burst arrives to the optical switching matrix or after some delay in an fiber delay line (FDL) buffer, the SCU will then select the FDL of the optical buffer and configure the optical switching matrix to let the data burst pass through.Otherwise, the data burst is dropped.If a data burst enters the optical switching matrix before its BHP has been processed (this phenomenon is called early burst arrivals), the burst is dropped.This is because data bursts are optical analog signals.If no path is set up when a data burst enters the optical switching matrix, it is lost.Since a BHP and its data burst are switched in the SCU and the optical switching matrix, respectively, the delay introduced by the input FDL should be properly engineered such that under the normal traffic condition data bursts are rarely dropped due to early arrivals.www.gjournals.org 2 An intrinsic feature of the OBS is the physical separation of transmission and switching of burst payloads and their headers, which helps to facilitate the electronic processing of headers at optical core routers and provide endto-end transparent optical paths for transporting burst payloads.The OBS network can be envisioned as two coupled overlay networks: a pure optical network transferring data bursts and a hybrid control network transferring BHPs.The control network (plane) is just a packet-switched network, which controls the routing of data bursts in the optical network based on the information carried in their BHPs.It is expected that the above separation will lead to a better synergy of both very mature electronic technologies and advanced optical technologies.
The offset time allows the core router to be buffer-less, avoiding then the requirement of optical memories, e.g.fiber delay lines, required on the contrary by optical packet switching.The control packet carries relevant forwarding information, as the next hop, the burst length and the offset time.It precedes the data burst by a basic offset time that is set to accommodate the non-zero electronic processing time inside the network and dynamically set up a wavelength path whenever large data flows are identified and need to traverse the network.Only the control packet is converted between optical and electronic domains, therefore is the only information delayed because of the conversion.The function of the burst offset-time depends on the design of optical core routers.For optical core routers using input FDLs (fiber delay lines) to delay the arrivals of data bursts to the optical switching matrix, thus allowing the SCU to have sufficient time to process their BHPs, the main function of the offset time is to resolve BHP contentions on outgoing control packets of optical core routers .For optical core routers without input FDLs, the offset time should also allow the SCU at each hop along the path to have enough time to process the BHP before its associated data burst arrives.In both cases, the traffic condition in the network should be taken into account in choosing the offset time.The burst offset-time could also be adjusted to support QoS (Qiao et al., 1999) and may play an important role in traffic scheduling/management for optical core routers without buffer or with buffer of very limited storage capacity (Turner 1999).
The most used reservation protocol in OBS network is Just-Enough-Time (JET) (Yoo, et al., 2000).JET is a delayed reservation protocol which allows reserving a wavelength channel just for the burst duration, starting at the predicted burst arrival time.If the reservation is successful, the control packet is forwarded to the next node; otherwise the correspondent burst is blocked.The use of this reservation protocol doesn't allow to fully utilizing the bandwidth; in every channel there is portion of unused bandwidth between bursts that have made a reservation, called void.In order to get a good utilization of the available resources, an efficient reservation process is required.To this end, effective scheduling algorithms have to be developed (Chen et al., 2004).The scheduling algorithm is implemented into the control unit as well.The scheduler is responsible of scheduling and switching data burst on an output data channel and also of transmitting the control packet.There is a scheduler for each data and control channel pair and each scheduler only needs to keep track of the busy/idle periods of a single outgoing data and control channel.It first reads the arrival time and the burst duration in the control packet and then, using a given scheduling algorithm, it searches for an idle output data channel.Several scheduling algorithms have been proposed for OBS routers (Xu, et al., 2003).
The focus of this paper is on the impact analysis on the performance factors of control plane (network) for OBS.The basic concept of OBS and related issues of OBS control plane are described.The importance of control-plane in OBS networks has been identified and analytical model with performance oriented techniques for OBS has been formulated for reducing its effect on overall OBS blocking probability and hence throughput.It is observed that by optimally selecting the combination of burst length and offset time in the proposed model, the control-plane impact to OBS system performance can be made negligible as compared to data-plane.

ISSUES RELATED TO CONTROL PLANE
To address the issues related to control plane is very important , as a modern network like OBS needs to be capable to be rapidly reconfigured not only by the demand of an operator but also by customer request with the aim to achieve an efficient use of bandwidth, low latency and high degree of transparency.The arrival process of burst control packets depends on the burst traffic load, the burst assembly algorithm, in particular on the distribution of both the payload and the control packet lengths, the number of control channels and the transmission rates in both control and data channels.The following are some of the factors that influence the control plane operation for OBS based system performance.
1. Control plane (network) should have the ability to create, reconfigure or modify and tear down connections rapidly according to the traffic variations.Thus, an adequate signaling scheme to perform an efficient end-to-end connection provisioning without long delays and resource reservation is crucial.2. Depending on the use of either the conventional OBS architecture or the offset time-emulated OBS or their hybrid solution the offset time may either vary or do not vary inside the network.As a result the delay budget of bursts entering the node is either variable or fixed.3. Insufficient estimation of the offset time may produce burst losses.Thus, an efficient management of the offset time is required to facilitate the inter-working between legacy and OBS networks by having a horizontal and vertical control of the optical network.4. When a control packet arrives at a core node, the decision of the scheduling algorithm must be fast enough in order to meet time constraint given by the offset time.Thus, an efficient scheduling algorithm does not have to be too complex, i.e., not is time consuming.5. Considerable functionalities should be provided by control plane i.e. a robust but simple routing method to forward the bursts with support of TE and QoS provisioning and to collect and distribute accurately the updated information of the network's resource availability in order to provide acceptable burst loss rates.6.The current burst assembly mechanisms such as timer-based and threshold-based (Ge. et al., 2000), lack flexibility to actual network traffic and increases assembly overhead and delay.The existing limit factors bring pressure to bear on core network and make the network performance deteriorate.Assembly time is inversely proportional to the offset time, that is, the longer the assembly time, the less time will be available for setting an offset at the burst queue.To ensure a well-built Control plane, efficient designing of burst assembly, contention resolution, QoS provisioning and protection and restoration mechanisms are required.7. The routing is a crucial functionality for optimal network performance.Good TE techniques and an adequate routing method are desirable to achieve that objective: how to have a simple but robust routing method that make the data burst safety reach the destination (low burst blocking) is still an open investigating area.a.Another important issue related with routing is topology and resource information dissemination.That is what parameters should be use to define the current state of the links and core nodes in order to help the edge nodes routing mechanisms to decide the best route to assign and how to flood these information with accuracy.Though, forward the BHP as a datagram would increase the necessary signaling and routing information carried in it leading to a high core node time processing while each core node would have to determine the next hop according to the current network status, which seems impractical in a highly dynamic network like OBS.Much more intelligent bandwidth utilization and a better burst block probability can be reached if the traffic is properly engineered by using explicit-routed paths.Labeled OBS (LOBS) suggested in (Zhang, et al., 2004) is the starting point to design an intelligent and efficient routing strategy.8.A simple node controller consists of a single processor unit with a buffer handling all the burst control packets in a centralized way.More advanced controllers use distributed, pipelined and parallelized operation onto multiple processors.These architectures speed-up the processing of control packets.The main functions performed by the controller processors are forwarding of burst control packets, resources reservation (with contention resolution and QoS functions) for incoming burst payloads and configuration of the switching matrix.These functions may be realized with algorithms of different complexity and performance.The algorithm implementation can be either memory-based, where the processing time depends on the seeking time in the memory map, or combinatorial, where the processing time is constant.Both selection and implementation of algorithms influence the service time distribution of the controller.The number of both node input/output ports and data wavelengths has an impact on the amount of burst control traffic arriving to the controller.Either simple first-in, first-out (FIFO) or more advanced Queuing disciplines can be utilized for ordering the burst control packets according to their offsets.

PROPOSED MODEL OF OBS CONTROL PLANE
The proposed model for the OBS control-plane explicitly accounts for the finite delay estimate of arriving headers.In order to alleviate an excessive variation in burst size, an efficient burst assembly scheme named as Adaptive-Threshold with Fixed Maximum Time Limitation Burst Assembly (ATH-FMTL) has been used in the proposed model, which uses optimal burst length threshold and fixed maximum time limitation as the condition for burst generation.The burst length thresholds are increased or decreased in case the burst queue size, at the time of burst generation, is larger than upper threshold or smaller than lower threshold respectively.The packets arrive at the corresponding port and service class assembly queue becomes operative.To classify the packets into the appropriate burst, the decision making is performed based on the fact that every packet has a delay tolerance that allows for flexibility during packet routing and on the assumption that no packet has a delay tolerance less that the amount of time it takes to route the packet through the OBS network, using the shortest route to it's destination.Each burst length is estimated at the end of p t (prediction time) according to the past burst length value and current arrival traffic.Edge node determines the variable burst assembly duration (VBAD) by estimating burst size with current or previous load.

−
.The following is the detailed working of the proposed model to have efficient use of bandwidth, low latency and high degree of transparency.
1.It is considered that a WDM link is having ' P 'channels with ' Q 'control channels and ' Q P − ' data channels ).
2. Assume that the data channel rate is ' R ' Gb/s and the control channel rate is ' r ' Gb/s.The maximum link utilization is . 3. As a data burst can be sent out on a data channel only if its BHP can be sent out on a control channel, there is a minimum requirement for the average data burst length in order to prevent congestion on control channels (Callegati et al., 1999).4. Let ' ab L ' be the average duration of a data burst and ' abhp L 'be the average duration of a BHP (It is also considered that both control and data channels are fully loaded).5.Under above conditions, the maximum average BHP transmission rate is abhp L Q BHPs per microsecond and the maximum average burst transmission rate is . Let γ denotes the mean arrival rate of headers to a core node.If the arriving traffic is the aggregate of a large number of independently generated traffic streams, the header-arrival process can be well-approximated by a Poisson process (Izal et al., 2002).It is also defined that ' header min δ ' is the minimum of the header transmission duration (on the control channel) and the duration required to process a single header in a core-node controller.Therefore, ' header min δ ' is equal to the inverse of the maximum rate at which headers can be serviced in a core-node processor (i.e., the rate at which headers exit the system during periods when the header-processing queue is not empty).7.In a constant-offset OBS architecture, the time required to process each header is independent of the state of the system, so ' δ ', the system can be modeled as an M/D/1 queue (Barakat et al., 2007).Using this model, the control-plane processor blocking probability ' CPP B ' can be calculated as follow : Poisson arrival process has been used to compute the data-plane blocking probability ' dp B '.For a system with random header arrivals, it is not generally possible to completely eliminate control-plane blocking.However, by proper designing and provisioning the network, the contribution of control-plane blocking is made negligibly to the overall blocking probability of the system.Using above, total probability ) ( Total B can be formulated as follows: 10.In the proposed model, the dynamic route selection is performed using the hop-based metric equation (Zhang et al., 2004), as it results in better performance in terms of loss, since minimal numbers of nodes are selected in a path, thereby reducing the probability of contention.11.A modified Horizon scheduling algorithm with minimum reordering effects (MHS-MOE) in OBS networks has been used in the proposed model (Garg et al., 2009).This algorithm runs much faster than Min-SV (Minimum Starting Void) and is significantly simpler than Min-SV in terms of complexity.

SIMULATION MODEL
NS-2 network simulator (NS-2, 2003) has been used in order to evaluate the performance of the proposed model of OBS control plane.The OBS core node architecture is equipped with N x N optical fibers capable of supporting K wavelengths each.This OBS core node mainly consists of input Fiber Delay Lines (FDLs), an optical switching system and a control unit.The input FDLs are used to delay incoming data bursts, thus allowing the control unit to have enough time to process the associated control packet.It is assumed that the optical switching matrix is equipped with a set of full range wavelength converters and that the system is buffer-less, i.e., no fiber delay lines are available in order to resolve contention for an output fiber, output wavelength.The control packets are processed by the control unit.The control packet is converted from optical to electrical domain in order to obtain information related to data burst.The control unit works as an electronic router and keeps the routing information.The network topology investigated in this paper is reported in Fig. 2. The reference network topology is composed of 9 nodes: four edge (source) nodes, one core node and four destination nodes.Edge nodes collect traffic from legacy interfaces and generate optical bursts adopting the assembly algorithm of the proposed model.It is also assumed the number of control channels is high enough to carry entire control traffic.The transmission bit rate of data channel is 40 Gbps.Switching time is 0.1µs.The delay introduced by the FDL is 0.30 of a mean burst length.It is assumed that a burst can only be buffered only once.Both burst length and inter-arrival time are exponentially distributed.The offset time is uniformly distributed over different ranges.The link load changes between 10 and 100% of maximum load (applies to edge nodes).The load at the core node will be slightly smaller as some bursts will be lost in the incoming links.A data-plane offered load per wavelength of 0.5, average burst length of 2Mb, assembly time (0.02s), normalized offset sizes ranging from 1 to 90, target blocking probability (1e-3) and a control-packet processing duration of 6 µs.All simulation results have 95% level of confidence (It is achieved by means of at least ten repetitions of the same simulation).
Fig. 2: Star-topology (Four edge nodes and two core nodes)

RESULTS
From fig. 3, it is observed that control-plane using proposed model has lower blocking probability with increasing offset duration.Since lower control-plane blocking results in a (slightly) higher data-plane offered load, data-plane blocking tends to increase with increasing offset duration.Therefore, the overall blocking rate decreases with increasing offset duration and thus minimal overall blocking is achieved with the proposed model.) present the impact of the network transmission and buffering capacities variation on the average burst rejection ratio.It is observed that the increase of the network buffering or transmission capacity decreases the average burst rejection ratio.The increase of the number of wavelengths per network link decreases the traffic load on each output channel, which decreases the contention probability.This will decrease a data burst traffic loss and so decrease the average burst rejection ratio.Moreover, the increase of the buffering capacity decreases the loss probability due to the lack of buffering units, which decreases data burst traffic loss; thus, the decrease of the rejection ratio.It is known that the control-plane load is directly proportional to the number of data wavelengths, offset duration and burst length in the system.In fig.9, it is observed that by using proportionate values of burst length, offset duration (assuming number of wavelengths ( 32)), control-plane blocking to OBS decreases and also target blocking probability 1e-3 (as formulated by proposed model) has been achieved.Thus, it is shown that control-plane blocking can be reduced by increasing burst length or offset duration.

CONCLUSIONS
In this paper, to ensure the reduction in impact of control plane in terms of packet loss-rate, latency, throughput etc to OBS system performance, a mathematical model of OBS control plane based on optimal assembly and scheduling schemes has been formulated for designing and provisioning of these networks.Using the proposed model, the tradeoff between the selection of optimum offset size and average burst length has been obtained for

Fig. 9 :
Fig. 9: Burst loss probability Vs Burst length (for different values of offset duration) Control packet is sent to OBS core network at timeτ ;τ = Average burst rejection Vs Buffering capacityFrom fig.7, it is observed that loss probability of control packets decreases for the system with different processor load [0.99, 0.9].