The intelligent inter-ring route controller (IIRC) is to determine a proper ringlet (CW or CCW) for an incoming inter-ring new call request at bridge. The determination of ringlet is based on the load balancing principle, in which the CW or CCW ringlet with lower congestion degree and higher service rate will be chosen. The congestion may come from the bridge node or the CW (CCW) downstream node. The former is related with the two STQ lengths of the associated interface in the bridge
node given in Figs. 3.1 and 3.2. Thus as shown in Fig. 3.3, the IIRC designs a fuzzy bridge-node congestion indicator (FBCI) to intelligently detect this congestion. The latter is related with the received fairRate from the downstream node of the associated ringlet. Therefore, the IIRC designs a PRNN (pipeline recurrent neural networks) downstream-node fairness predictor (PDFP) to predict the CW or CCW downstream-node congestion degree. Finally, the IIRC designs a fuzzy route controller (FRC) to determine a proper ringlet for the incoming inter-ring new call request. It receives the congestion indication from FBCI, denoted by CI , and the predicted mean received fairRate from PDFP, denoted byRf , as input linguistic variables. Also, it considers the service rate of the CW or CCW ringlet at the bridge node, denoted by R, and the number of hops between the bridge and the destination, denoted by H, as input linguistic variables. Notice that the ringlet service rate at the bridge node is related with the received fairRate and more hops consume more system bandwidth. The FRC calculates the preference value of route, denoted by Pv, for CW and CCW interfaces and selects the ringlet with larger Pv as the proper ringlet route for the incoming inter-ring new call request.
Figure 3.3: Intelligent inter-ring route controller (IIRC)
Fuzzy Bridge-Node Congestion Indicator (FBCI)
The fuzzy bridge-node congestion indicator (FBCI) considers four measures as the input linguistic variables to determine the congestion degree of the bridge node at the CW or CCW interface. They are STQ lengths in the ingress buffer and the ringlet buffer, denoted by QSI and QSR, respectively, the amount of the reserved bandwidth for
high class traffic (which are stored in PTQ), denoted by BA, and the equivalent capacity of the incoming inter-ring new call, denoted by Ec. Note that the equivalent capacity for a new call can be estimated from its traffic description parameters: the peak rate, mean rate, and peak rate duration of packets. Among the four measures, the two STQ lengths are the more essential measures to indicate the degree of the congestion in the RPR bridge node. The BA occupancy is highly correlated with the STQ due to the fact that the system bandwidth is allocated to high priority traffic first.
Also, the amount of Ec can cause the increment of the STQ length. The output linguistic variable of the FBCI is the congestion degree of the CW or CCW interface of the bridge, denoted by CI.
Term sets for the four input linguistic variables and the output linguistic variable are defined as T(QSI(QSR)) = {Short (S), Medium (M), Long (L)}; T(BA) = {Few (Fw), Many (Ma)}; T(Ec) = {Small (S), Large (L)}, and T(CI) = {Very Low (V L), Low (L), Medium (M), High (H), Very High (VH)}.
Here, the triangular function f
x:x0 ,a0 ,a1
and the trapezoidal function
x:x0 ,x1 ,a0 ,a1
g are used to define the membership functions for the terms in the term set. These two functions are
0 0 0 0 triangular or the trapezoidal function.
Membership functions for S, M and L in T(QSI) are expressed as
)
threshold in percentage. Note that if the STQ length is larger than lth · Ls, the bridge is in congestion and the fairness algorithm will be enabled. Membership functions for S, M and L in T(QSR) are similar to S, M and L in T(QSI), respectively. Membership
functions for Fw and Ma in T(BA) are defined
as Fw(BA)g(BA ;0,0.025C ,0,0.025C) and Ma(BA)g(BA ;0.1C,C ,0.075C,0) , where C is the link capacity. Membership functions of terms in T(Ec), are defined as
)
and Rvideo are the minimum demand of the mean rates of the voice and video traffics, respectively, and Rh is the maximum demand of the mean rate of the video traffic to provide the high quality video. Membership functions for terms of output linguistic variable CI are defined as VL(CI) f(CI ;0.1,0,0) , L(CI) f(CI ;0.3,0,0) ,
Table 3.1: The Fuzzy Rule base of FBCI
As shown in Table 3.1, there are 24 fuzzy rules for FBCI, where the notation ”X”
in this table represents ”don’t care” of the linguistic variable. The order of significance of the input linguistic variables for the FBCI would be QSI, QSR, BA, and Ec in sequence. The bridge will be in high degree of congestion if its two STQ queue lengths are close to or longer than the threshold (the corresponding terms of QSI and QSR, are Medium or Long). Finally, FBCI adopts the max-min method for fuzzy inference. The defuzzification method adopted is the center of area defuzzification
method.
PRNN Downstream-Node Fairness Predictor (PDFP)
The bridge uses the received fairRate from associated ringlet of downstream node to discern the congestion degree of the downstream node. If the received fairRate is high, it means that the downstream nodes’s STQ can accept more flows and the bridge can raise its service rate. Otherwise, it means that the downstream nodes’s STQ is going to be full or has overflowed and the bridge should decrease its service rate. However, by the AM fairness algorithm considered here, thereceived fairRate would vary. This high variation of the received fairRate would make the bridge not easily detect if its downstream node is in congestion or not. Therefore, we originally choose an average received fairRate over the past m periods from the current nth period, denoted by Rf(n), as the input variable, where m is the size of the observation window, m ≥ 1. The Rf(n) could be appropriate to detect the congestion situation of the downstream nodes during a period and it is expressed by
m
where Rf (n) is the received fairRate at time n. Also, since the bridge node routes the traffic flows call by call, the next-step mean received fairRate could be more appropriate to determine the route for an accepted new call. Here, a pipeline recurrent neural networks (PRNN) is adopted to design the PRNN downstream-node fairness predictor (PDFP). The fairRate with one-step prediction as a function of p received fairRates and q previously predicted fairRate, denoted by Rf(n1) or Rf for and H(·) is an unknown nonlinear function to be determined. The pipeline recurrent neural network (PRNN) prediction is a fast, low-complexity, and non-linear one that can approximate the function H(·).
The incremental change of synaptic weights is according to the steepest decent method. Also, the training of PRNN consists of two stages. During the off-line training phase, the PRNN, fed with the received fairRates, adjusts the synaptic weights recursively until the root mean square error (RMSE) of the desired prediction output is lower than the criteria. During the on-line training phase, the PRNN fairness
predictor obtains the fairRate predictions at (n+1)th period, Rf(n1), from the output of the first neuron of the first module, and receives the new fairRate Rf(n1); then it adjusts the synaptic weights using the real time recurrent learning (RTRL) algorithm. Due to the on-line learning capability, PDFP can adapt its wights to the current load conditions other than those set in the off-line training phase. If a PRNN contains q modules and M neurons per module, the computational complexity would be O(qM4). However, when the system is in operation and the PRNN has determined each parameter by learning, the computational complexity is reduced to O(1).
Fuzzy Route Controller (FRC)
The fuzzy route controller (FRC) is to determine the route preference values, Pvs, for both of CW and CCW ringlets. The determination is based on four input linguistic variables of ringlet: the congestion indication of the bridge node, CI , the predicted mean received fairRate, Rf , the current service rate of the ringlet, R, and the number of hops to destination, H. The higher value Pv of a ringlet means that the ringlet is more suitable to accept the incoming new call request. Term sets for the input and output linguistic variables are defined as T(CI) = {Low (Lo), Medium (Me), High (Hi)}, T(Rf ) = {Small (Sm), Medium (Me), Large (La)}, T(R) = {Low (Lo), High (Hi)}, T(H) = {Few (Fw), Many (Ma)}, and T(Pv) = {Unsuitable (U), Weakly Unsuitable (WU), Weakly Suitable (WS), Suitable (S)}. Membership functions for terms of Lo,Me, and Hi in T(CI) are defined as Lo(CI)g(CI ;0,0.25 ,0,0.25),
, where v denotes the unreserved bandwidth for the low priority traffic at the bridge node and v = C − BA.
Similarly, membership functions for T(R) are defined as ) the total capacity of the fiber link. Membership functions for terms of Fw and Ma in T(H) are defined as Fw(H)g(H ;0,N/3,0,N/3) and
network. Finally, membership functions for terms in T(Pv) are defined as
0) represents ”don’t care” of the linguistic variable. The rules are designed according to the load balancing principle for FRC, and the order of significance of the input linguistic variables for the FRC is CI , Rf , R, and H. The low congestion degree of ringlet interface (CI = Lo) and the large or medium predicted mean received fairRate (Rf = La or Me) would make the inter-ring new call have more chance to enter the interface. However, the low congestion degree of ringlet interface (CI = Lo), but the small predicted mean received fairRate (Rf = Sm) which means that the downstream nodes may incur congestion, and the high ringlet service rate (R = Hi) would make the variable of the number of hops to destination H significant. If H is Few, the new call will be weakly suitable for the ringlet, while if H is Many, the new call will be weakly unsuitable for the ringlet. On the other hand, the high congestion degree of ringlet interface (CI = Hi) and the small predicted mean received fairRate (Rf = Sm) would make the inter-ring new call have less chance to enter the interface. However, the high congestion degree of ringlet interface (CI = Hi), but the large predicted mean received fairRate (Rf = La) which means that the downstream nodes are free of congestion, and the high ringlet service rate (R = Hi) would similarly make the variable of the number of hops to destination H significant. The fuzzy inference algorithm also adopts the max-min inference method, and the defuzzification method, the center of area defuzzification method.
Simulation Result
Simulations are here conducted to compare the performances of proposed IIRC, and SPRC. Also, an intuitive queue-length threshold route controller (QTRC) is included, which determines a proper ringlet depending on the shorter STQ length of ingress buffer. Traffic flows from R1 to R0 at the bridge node are considered.
Referring to Fig. 3.1, assume that there are M = 16 non-bridge nodes on R0, the link capacity is C = 10.0 Gbps, and sizes of the two PTQs and the two STQs are 40 Mbyte with threshold lth = 1/4. Three kinds of calls are considered in the system: voice, video, and data. The two-state Markov chain is used to model packet traffic flow of calls with two different arrival rates and two state transition rates. Then the peak rate
Rp, the mean rate Rm, and the mean burst period Tp with the four previous rates can be obtained.
For voice packet generation process, during the ON (talkspurt) state, voice packets are generated with rate 21×10−4; during the OFF (silence) state, no packets are generated. A voice source has two transition rates of 4 × 10−5 and 8 × 10−5 in the ON and OFF states, respectively. The packet size is fixed at 70 bytes, and thus the generation rate is constant bit rate (CBR) during ON state. The arrival process of a voice source was assumed that Rp = 21 × 10−4, Rm = 7 × 10−4, and Tp = 1.3s. Two kinds of video packet generation processes are assumed: the intraframe and interframe generation processes. The intraframe (I-frame) generation process is similar to the voice packet generation process with generating rate 5×10−2, and two transition rates of 4 × 10−5 and 8 × 10−5 in the ON and OFF states, respectively. The arrival process of the I-frame of video packet source was assumed that Rp = 5 × 10−2, Rm = 1 × 10−2, and Tp = 0.1s. The interframe (B- and P-frames) generation process includes B-frame-bit-rate and P-frame-bit-rate video services, and their generation was characterized by Bernoulli processes with rates θB and θP , respectively. For B-frame-bit-rate of the B-frame of video packet source, it was assumed that Rp = 2 × 10−2, Rm = 2 × 10−3, and Tp = 0.01s, which is given θB = 0.1; for P-frame-bit-rate of the P-frame of video packet source, it was assumed that Rp = 1×10−2, Rm = 2×10−4, and Tp = 0.01s, which given θP = 0.02. The I-frame packet size is fixed at 1000 bytes, and the generation rate is CBR; the B-frame, and P-frame packet sizes are uniformly distributed over 100 and 1518 bytes and the generation rates are with generation of variable bit rate (VBR). The data packet generation process includes high-bit-rate and low-bit-rate data services, and the generation of high-bit-rate data packets and low-bit-rate data packets are characterized by Bernoulli processes with rates θ1 and θ2, respectively. For highbit- rate of data source, it was assumed that Rp = 7 × 10−2, Rm = 7 × 10−3, and Tp = 0.03s, which is given θ1 = 0.1; for low-bit-rate of data source, it was assumed that Rp = 3.5 × 10−2, Rm = 7 × 10−4, and Tp = 0.03s, which is given θ2 = 0.02. The data packet sizes are uniformly distributed over 100 and 1518 bytes and the generation rates are with generation of variable bit rate (VBR). The parameters, Rvoice, Rvideo, and Rh, are set to 64kbps, 640kbps, and 6.4Mbps, respectively.
Fig. 3.4(a), (b), and (c) show the average packet dropping probability, the average packet delay, and the throughput, respectively, for the proposed IIRC, QTRC, and SPRC, versus the traffic intensity from the R1 to R0 at the bridge in a balanced scenario. The traffic intensity at the bridge is here defined as the total arrival packet rate over the capacity of the fiber link. In this balanced scenario, in R0, both the local CW ringlet traffic intensity from node 16 to bridge and the local CCW ringlet traffic
intensity from node 1 to bridge are fixed at 0.6, and the varying inter-ring traffic intensity is from 0.3 to 0.7; the add traffic intensity of node 1 in CW ringlet and the add traffic intensity of node 16 in CCW ringlet are both fixed at 0.2; the probability of the destination of the incoming new calls is uniformly distributed over nodes on R0. It is found that the packet dropping probability and the average packet delay of both CW ringlet and CCW ringlet are almost the same for IIRC, QTRC, and SPRC. The results show that IIRC, QTRC and SPRC can achieve the load balancing in this balanced scenario. It is because the probability of the destination for the new call request is uniformly distributed over nodes; the routing policy of QTRC is simply according to a shorter STQ length of ingress buffers and the routing policy of SPRC is based on the shortest path. Also, this justifies that the IIRC, which chooses a suitable ringlet with lower congestion degree and higher service rate, is well designed. Furthermore, IIRC has the lower packet dropping probability by about 16% and 29%, the smaller average packet delay by about 9% and 21%, and the higher throughput by 5.1% and 7% in heavy bridge traffic intensity than QTRC and SPRC, respectively. It is because QTRC does not consider the number of hops to destination, and thus QTRC would route calls to pathes with more nodes and then consume more bandwidth. Also, in the situation that many incoming new calls just happen to have the same destinations, SPRC’s routing policy would make the STQ overflow. However, IIRC decides a suitable route for each call independently based on congestion degree and service rate.
Fig. 3.5(a), (b), and (c) show the average packet dropping probability, the average packet delay, and the throughput, respectively, versus the bridge traffic intensity in an unbalanced scenario. Here, the probability of destination of nodes for new calls is non-uniformly distributed, where node 1 (9) to node 8 (16) are with the same probability 1/40 (1/10). It can be found that the packet dropping probabilities and the average packet delays of CW and CCW ringlet by IIRC and QTRC are still almost the same, while these by SPRC are quite different. We can deduce that the IIRC can indeed perceive the congestion degree of CW and CCW ringlets and sophisticatedly achieve the load balancing by overall considering the congestion degree, the received fairRate, the ringlet service rate, and the number of hops to destination. QTRC could avoid enlarging a longer STQ length of the ingress buffer due to its routing policy. Moreover, IIRC improves by about 10% and 220% in packet dropping probability, and by about 13% and 18% in average packet delay, by about 6% and 19% in throughput in heavy traffic intensity over QTRC and SPRC, respectively. It is because the SPRC scheme would route most calls via the CCW ringlet for most destinations of incoming new calls that are on the up side of the bridge. This will make the STQ occupancy of CCW interface in R0 exceed a threshold and thus SPRC gets a worse throughput.
Figure 3.4: The performance comparison for IIRC, SPRC, and QTRC in the balanced scenario
Figure 3.5: The performance comparison for IIRC, SPRC, and QTRC in the unbalanced scenario
Fig. 3.6 shows the bridge throughputs under IIRC, IIRC without considering EC and/or BA in a balanced scenario as given in Fig. 4.7 It is found that the IIRC has the largest throughput; it improves by about 1.5%, 3.6%, and 6.7% over IIRC without considering EC, IIRC without considering BA, and IIRC without considering BA and EC, respectively. These can justify that the input linguistic variables BA and EC are essential, and the BA input linguistic variable is more important than EC.
Figure 3.6: The comparison of bridge throughputs under different schemes versus the inter-ring traffic intensity
IV. Reservation Slotted OBS Rings with Wavelength Assignment and Traffic Control
System Model
Assume that the slotted OBS ring is with N nodes and is constructed by two unidirectional, counter-rotating ringlets, named ringlet-0 and ringlet-1. Each node has two pairs of input and output ports to communicate with neighbor nodes. Node X (Y) is said to be an upstream (downstream) node of node Y (X) on ringlet-0 or ringlet-1 if the node Y (X) traffic becomes the received traffic of node X (Y) on the referenced ringlet. There are four classes of service considered: voice, video, the hypertext transfer protocol (HTTP), and the file transfer protocol (FTP).
Slotted OBS Rings
The slotted OBS ring contains W data wavelengths, denoted by , …, 1 W, and one control wavelength, denoted byc, on each ringlet. As shown in Fig 4.1, there is a frame structure on each wavelength; each frame of data wavelength is composed of S slots; and there is a wavelength reservation (WR) transited each frame in control wavelength and rotated around the ringlet for nodes to make wavelength reservation. The WR comprises four kinds of messages of the next frame: total number of available free time slots, denoted by E, the advertised fair rate, denoted by Fa, the status of slots in the ith wavelength, denoted by Si,
,
1iW and the CBj relating to the DBj, 1 j J, where J is the number of the transit DBs at the next frame. Note that Fa is used to avoid the congestion and to achieve the fairness between each node by limiting the amount of the ingress traffic into ringlet. It will be transmitted to upstream nodes. The Si contains S bits, and each bit in Si represents the status of the corresponding slot with bit 0 (1) to indicate free (busy) in the wavelength i at the next frame. The CBj comprises information of DBj such as its destination node, source node, wavelength number, start slot position, burst length in unit of slots, and service class.
Node Architecture
As shown in Fig. 4.2, each node i is in the slotted ring, 1iN, contains
As shown in Fig. 4.2, each node i is in the slotted ring, 1iN, contains