Huifeng Bai, Dongshan Wang, Yanbin Song

Beijing SmartChip Microelectronics Company Limited, Beijing 102200, China

* The corresponding author, email: lancer101@163.com

I. INTRODUCTION

As multiple services emerge constantly with increasing demand on network capability,flexibility and reliability, optical access network technologies have achieved great development and the bandwidth of PON (Passive Optical Network) evolved from one Giga-bite up to even 40-Giga-bite [1-2]. To better the quality of service (QoS) with high efficiency, the High-Speed PON must be active to support diversified services, which are full of complexity, difficulty and dynamic. Thus,the High-Speed PON must be able to be fully aware of characteristics and requirements of multiple services and to match these services with outstanding performance. This ability can be called “Services Awareness”.

In fact, the services awareness contains three aspects, which includes awareness objects, awareness methods and awareness results. Therefore, this services awareness has already become an important issue in many fields, including network design, network management, and network security, etc [3-5]. Great efforts have been made on analyzing and classifying technologies of services.Deep research on relation between statistic characteristics and service protocols has been done in [6-7]. However, those approaches are not so accurate for service-layer protocol.As furthering research, series researches on data-traffic identification and classification mechanism have been reported in [6-8], which is based on the packet length, arrival time interval and arrival sequence. As one solution, a Bayesian classifier based services-awareness(BC-SA) has been presented in Ref. [8]. This also implies that more pattern recognition algorithms also have great potential to realize service-awareness. However, there exist several problems in current neural network algorithms, including consuming time and computational complexity and so on. The echo state network (ESN) [9] is a strong candidate to better service-awareness performance of highspeed PONs, as ESN has great advantages to enhance reliability, to improve accuracy [10-11].

In this article, the author proposed a Simplified Echo State Network based services awareness(SESN-SA) mechanism in 10G-EPON.

However, the complexity of original ESN makes it not so feasible for terminal netword elements to be in full charge of ESN training and computation. Another challenge comes from the real-time requirement by high-speed PON itself, as the software processing of original ESN will results in unnecessary time-delay. Additionally, considering the architecture of PON system, there still exist great problem to make the ESN be suitable to the high-speed PON system, which is composed by optical line terminal (OLT) and a number of optical network units (ONUs). The OLT-ONUs architecture also makes it unfeasible to allocate the whole ESN only in OLT, since ONUs are hard to benefit from original ESN based services-awareness. However, several challenges are still there to realized services awareness in high-speed PON system. Due to the architecture of PON system, each ONU must conduct independently services awareness, since the ONU is the terminal device of services packets access. Therefore, it is of ultra importance to make sure the consistency of services awareness among ONUs.

With the aim to solve this problem, this article proposes an simplified ESN based services awareness (SESN-SA) scheme for High-Speed PON. In this proposed approach,the so-called “SESN Master” module is embedded and running in the OLT with the whole function of SESN, while the “SESN Agent”works with sampled functions in each ONU.The SESN Master is responsible of initiating and training the SESN to form the reservoir and providing the reservoir information to all ONU through broadcast. And each SESN Agent directly uses this reservoir information from OLT to form its SESN and conduct independently service-awareness operation.

II. SIMPLIFIED ECHO-STATE-NETWORK BASED CLASSIFICATION

Traditionally, the echo state network (ESN)is composed by three layers, including input layer, reservoir and output layer [9-13].Moreover, each layer has a number of units in itself and the basic theory of ESN is given in formula (1) and formula (2). In general, there are three types of connection weight matrix in the ESN: Win between the input layer and the reservoir, W of the reservoir and Wout of the output layer. Additionally, u(n), x(n) and y(n) represents individually the input variable,state variable and output variable. Within the reservoir, there exist a great number of units and the topology to represent relations among these units. In fact, the topology is one of core issues which directly determine the performance of ESN. However, the highly complex topology in traditional ESN has brought a great challenge to realize the ESN in the manner of hardware, which was traditionally realized with software and failed to meet real-time demands of services.

With the aim to make the ESN suitable for its hardware application in high-speed PON system, some important simplification and optimization of topology in reservoir are necessary to reduce the complixity of ESN.According to research results of Ref. [14-15], simple time-delay system with nonlinear nodes is able to deal with information processing tasks of complex system, which could satisfy the requirement to build the reservoir of ESN. On this basis, a Simplified-ESN (SESN)is proposed with less non-linear units and simple topology in its reservoir, as shown in Fig.1. In this SESN, the reservoir only includes N nonlinear units to form a ring-topology, and this makes the SESN simple enough for it to be realized by hardware. Thus, there are two key issues to construct the SESN: one is the time-delay formula to generate these non-linear units in the reservoir; the other one is the information procession of the ring-topology.

To generate these non-linear unitsx(i)of the reservoir, the system dynamics equation is introduced, which must be characterized by fully dynamic feature as defined in (3).

Furthermore, the Nicholson’s Blowflies equation is considered as depicted by formula(4), since it satisfies the demand of the dynamic feature.

Where,Pis the average generation rate,and1/ɑis the total number with the maximum generation rate and?is the death rate. After discretization procession, the formula (4) can be transformed into formula (5). Using this formula (5), each unit in the reservoir can be generated.

Fig. 1 Simplified ESN

According to the ring-topology feature, the W in the reservoir can be obtained by formula(1), where the value of wij is defined as formula (6).

The main idea of SESN classification is given in formula (8), which means input signals must be kept without change until state variables are quite stable. This method still keeps the advantage of ESN in which the training is fast and simple, and shows great potential in classification.

On this basis, the complexity of original ESN is greatly reduced, which allows this proposed simplified-ESN to be realized by hardware instead of software and to improve the processing-delay.

III. SIMPLIFIED-ESN BASED SERVICES AWARENESS IN HIGH-SPEED PON

3.1 Basic principle

In fact, there exist several key parameters can be used comprehensively to identify the type of service, including packet size, inter-arrival time and service-duration time. Those parameters of service characteristics can be gained through independently different methods. By making full use of these parameters as the input information, the Simplified-ESN is able to perform services-awareness function in highspeed PON. Moreover, types of services are also curtained in this article. And QoS of these multi-services can be mainly divided into three categories: packet-loss-rate, time delay and jitter, where each category is with three classes:high-level, middle-level and low-level.

As is depicted in Fig. 2, a Simplified-ESN based service awareness (SESN-SA) mechanism is proposed as one feasible approach. By making full advantage of the Simplified-ESN,this proposed SESN-SA can be realized directily by hardware. Thus, it will greatly better efficiency than the original ESN based service-awareness, which is realized by software.

In fact, this proposed approach can be deployed in various kinds of PON systems. In this paper, the 10G-EPON is chosed as tipycal high-speed PON to perform the services awareness. According to the architecture of 10G-EPON system, the services awareness is realized by allowing each ONU to independently perform Simplified-ESN based services awareness with low complexity and by centralized Simplified-ESN training in the OLT. the SESN Master module is responsible for providing ONUs with well-trained Simplified-ESN from the view of the whole PON system, which implies that the SESN Master is a complex one. This SESN Master must complete the initiation and training of the SESN and provide this completed SESN to all ONUs. Thus, the Agent ESN in each ONU just only needs to obtain ESN and to directly perform service-awareness using this SESN. It will abstract service-characteristics parameters from newly arriving services and to compute classification results for ONU to conduct schedule operation.

3.2 Pre-processing of parameters data

Aimed to avoid the over fitting phenomenon of the ESN or SESN, the pre-processing of parameters data is necessary, before SESN training and classification operation. And these parameters of serviceiincludes: packet sizePSIZE(i), inter-arrival timePINTERVAL(i)and service-duration timePDUR(i). Therefore, the pre-processing can be conducted using the formula (9).

Fig. 2 Architecture of the SESN-SA system

Where, thePSIZE_MAXis the maximum value of packet size, and thePINTERVAL_MAXis the maximum one of inter-arrival time andPDUR_MAXis the maximum service-duration time among all kinds of services.

3.3 Simplified-ESN training in OLT

In this proposed mechanism, the SESN initiation and training are centralized and conducted in the OLT, in order to make sure the consistency of services awareness among ONUs,since the OLT has the whole view of all services in the 10G-EPON system.

Before the SESN classification training begins, all parameters of SESN must be well set by the Main SESN, including the number N of units in reservoir, the internal connection weights matrix Win, etc, and the inter-connection weight matrix W is generated. After this training is fully completed, the information of the well-prepared reservoir is distributed to all ONUs. The detailed procedure of SESN initiation and training in OLT is depicted as follows:

Step 1: Set the processing units numberN,the number of output units and other necessary parameters, whereN=200,P=19.8,ɑ=1,?=0.8,θ=0.2andx(0)=5according to Ref.[16];

Step 2: Compute values of all nodes using the formula (5);

Step 3: distribute these nodes to form the ring topology and determine whether the connection between them exists according to the ring topology in the reservoir;

Step 4: Initiate the input matrixWinand generate the internal connection weight matrixWusing the method mentioned above in formula (6-7);

Step 5: Input all training samples into the SESN that has already been initiated fully;

Step 6: update the output matrixWoutusing the the pseudo inverse algorithm.

Step 7: The SESN Master completes the SESN training.

Through the initiation and the training procedures, the SESN Master is well-prepared in OLT. Later, the detailed information of this SESN Master will be broadcast by the OLT to all ONUs. Each ONU will draw this information to form their SESN Agents and conduct SESN based distributed services awareness independently.

3.4 Simplified-ESN based servicesawareness in the ONU

Generally, the services scheduling of 10G-EPON system is realized under the combination of the dynamic bandwidth algorithm(DBA) and intra-ONU scheduler [17-20]. In this proposed simplified-ESN based services awareness (SESN-SA) mechanism, ONUs directly play key roles to be aware of decertified services and to perform schedule function according to awareness results.

For implement, the SESN Agent classification module is embedded into each ONU and its task is to perform multi-services awareness together with the scheduler in ONU under the control of DBA in 10G-EPON. On receiving the detailed information of the SESN Master from OLT, each ONU immediately conducts its multi-services awareness initiation and constructs the agent SESN module the same with the well-trained main SESN according to the detailed information. After all ONUs finish this initiation, the ESN based multi-services awareness start to work. And the scheduling process in each ONU is also depicted in Fig.2.

In Fig. 2, services packets arrive with different priorities, and their characteristics parameters (including packet sizePSIZE(i), inter-arrival timePINTERVAL(i)and service-duration timePDUR(i)) are abstracted and transformed intou(n)set of the input-layer of SESN for classification computation. And according to output results of the SESN, these services are classified by the SESN Agent module and put into corresponding queues. Then, the scheduler of ONU will move some packets according their priorities into the buffer which would send these packets to the OLT.

The complexity of SESN classification is one of key issues, which must be low enough for implement. According to the SESN model,it only need two times of multiply operation to deal with one kind of classification parameter,where the complexity isO(1). If the number of classification parameters reachesm, the complexity is justO(m)×O(1)for the SESN classification computation in each ONU.

IV. TESTING RESULTS AND ANALYSIS

In order to evaluate the performance of the proposed scheme, a testing platform of 10G-EPON system with 32 ONUs is build.The comparison is made between the original ESN based service-awareness (ESN-SA) and the proposed SESN-SA, where the ESN-SA is realized by software and the SESN-SA is directly by hardware using FPGA (Field-Programmable Gate Array). The “SESN Agent”module is running in each ONU, while the“SESN Master” module works in the OLT.

In this test, three classes of services are set to produce traffic load, and requirements of services to network are represented by three parameters: the average packet delay, the packet loss rate and the jitter. Moreover, each kind of services is divided into three levels in term of priority: class1 with the highest level,class2 with mediate level and class3 with the lowest level. Additionally, the class_1 service takes the proprotion of 30%, and the class_2 service are set to be 20% respectively, and the class3 service is 50%. The traffic load of each class service follows the Passion distribution.Addtionally, at least 8 types of services is are used for testing, as is given in Tab. 1. The Ethernet Analyzer MD1230B is used to generate traffic load that covers these services.Moreover, various traffic load characteristics can be edited to produce services traffics with corresponding features, using the Traffic Generation function.

Fig. 2 shows the accuracy rate result between the original ESN and the simplified ESN. There are two obvious observations can be gained. One is the relation between accuracy rate and training times. As the training times increase hundred by hundred, the accuracy rate of prediction soars obviously. After enough training, the accuracy rate shows the trend toward 100%. The other one is that the accuracy rate of SESN is finally close to the original ESN, after 2500 training times. That means the SESN can reach the same accuracy with the original ESN, because it benefits greatly from introduction of system dynamics equation with full dynamic feature.

As to the cross-check result, each packet will be marked with its value of classification result using the TOS (Type Of Service) in the packet, for convince to check the accuracy of SESN classification result later. Moreover,classification accuracy can be calculated by checking the TOS value of each service. In this paper, first 3 bits of TOS is used to mark services-awareness results for different kinds of traffics, where each kind of service after classification is given a priority value from the set {001,010,...,110,111}. After the SESN is well trained, the final service-awareness result can be obtained as shown in Tab. 2.

For convenience, the Weighted Round Robin (WRR) based DBA is adopted by this testing system as the basic QoS schedule approach. At least, there are three types of queues with corresponding weighted value in the ONU, and scheduling performs among three queues in orders to make sure that curtained service-time can be achieved among different queues.

Fig. 3 Comparison of accuracy rate

Table I Services types in test

Table II Service-Awareness Results of SESN in Detail

The comparison of traffic load is depicted in Fig. 4, which includes bandwidth utilization rates of three kinds of approaches: the SESNSA, the ESN-SA and the original one without services-awareness, under the same traffic load.

Fig. 4 Comparison of bandwidth utilization

Fig. 5 Comparison of time delay

Fig. 6 Comparison of packet loss rate

Testing results about services awareness performance are compared from aspects of time delay, packet loss rate and jitter. These comparisons results are given from Fig. 5 to Fig. 7.

The delay time of delay-sensitive service is given in Fig. 5. Obviously, delay time of all services soars as the traffic load increase.Under the same traffic load condition, class_1 service and class_2 service with SESN-SA show better value than their corresponding ones with ESN-SA, while class_3 service with SESN-SA gets the worst value. Because the class3 service has lowest requirement on delay time, it is tolerant to this time-delay performance. Additionally, the class1 services in SESN-SA can achieve the best performance.This result obviously suggests the advantag of the SESN-SA scheme in term of processing speed, because the SESN Agent works in ONU by hardware.

Fig. 6 shows the comparison of packet loss rate for packet-loss sensitive services,and these services are also divided into three levels (class 1, class 2 and class 3). Overall,the packet loss rate of all services soars as the traffic load becomes heavy. Similar to the comparison result of network delay time,class1 service and class2 service with SESNSA show lower packet loss rate than ESN-SA.And the class3 service with SESN-SA still gets the worst result, since it has the lowest requirement on packet loss rate. Thought comparison, the packet-loss-rate of SESN-SA is much more reasonable than ESN-SA. With the SESN-SA scheme, differences among those services are more obvious. Thus, the requirements of different classes of services on packet loss rate can all be further satisfied by using the SESN-SA mechanism.

Comparison result in Fig. 7 shows that the ESN-SA and the SESN-SA has similar performance in term of jitter. Generally, the jitter is caused mainly by the limited queue or buffer.Benefitting from greatly stronger hardware of the 10G-EPON system, the difference between ESN-SA and SESN-SA is not so obvious on jitter performance.

Combining all comparison results from Fig.5 to Fig. 7, the SESN-SA is able to achieve better performances comprehensively, especially in the time-delay performance. And the SESN-SA approach is able to enhance matching degree between services demands and 10G-EPON ability, with better time delay and more reasonable packet loss rate. Benefitting from the simplified inner topology of ESN, the SESN is able to achieve faster processing speed than the ESN. Thus, SESN can better QoS performance on the delay-time and packet loss rate at some degree. Therefore,the SESN-SA is suitable for the architecture of 10G-EPON, since ONU is the right one to face multiple services and to perform services-awareness directly with higher efficiency.

V. CONCLUSIONS

As ever-increasing multi-services had brought great challenge with demand of high matching degree between services and optical access networks, the high-speed PON was required to actively be aware of the characteristics of those services and to provide support with not only greater capacity but also better efficiency. In order to cope with this challenge, this article had proposed an Simplified Echo State Network based services awareness (SESNSA) mechanism in 10G-EPON, which presented firstly the simplified state network model and made full use of advantages of SESN to achieve the sercice-awareness function in 10G-EPON. With this SESN-SA scheme, better matching degree between multi-service and 10G-EPON system was able to be conducted.Simulation results showed that the SESN-SA scheme was able to match requirements by multiple services with better performance in terms of time-delay and packet loss rate.

ACKNOWLEDGEMENT

This work is supported by the Science and Technology Project of State Grid Corporation of China: “Research on the Communication Architecture and Hardware-In-the-Loop Simulation of Real-Time Wide-Area Stability Control for Electric Power System”.

Fig. 7 Comparison of jitter

[1] Zhang J, Ansari N. Toward energy-efficient 1G-EPON and 10G-EPON with sleep-aware MAC control and scheduling[J]. IEEE Communications Magazine, 2011, 49(2): 33-38.

[2] Dinh N, Walid A. Power saving protocol for 10G- EPON systems: A proposal and performance evaluations[C]// Proceedings of IEEE Global Communications Conference. IEEE, 2012.

[3] Mukai H, Tano F, Nakagawa J. Energy Efficient 10G-EPON system[C]// Optical Fiber Communication Conference. 2013:1-3.

[4] Elbers J.P. Optical access solutions beyond 10G-EPON/XG-PON, proceedings of 2010 Conference on (OFC/NFOEC) National Fiber Optic Engineers Conference, pp 1-3, 2010

[5] Roy R, Kramer G, Hajduczenia M. Performance of 10G-EPON, IEEE Communications Magazine,2011, 49(11): 78-85. 2011

[6] Dainotti A, Pescape A, Claffy K. Issues and future directions in traffic classification [J]. IEEE Network, 2012, 26(1):35-40.

[7] Zhang J, Xiang Y, Wang Y, et al. Network Traffic Classification Using Correlation Information[J].IEEE Transactions on Parallel & Distributed Systems, 2013, 24(1):104-117.

[8] BAI Hui-Feng, LI Ming-Wei, WANG Dong-Shan.Bayesian Classifier Based Service-aware Mechanism in 10G-EPON for Smart Power Grid[J].Acta Photonica Sinica, 2013, 42(6): 668-673.

[9] Rodan A, Tino P. Minimum complexity echo state network.[J]. IEEE Transactions on Neural Networks, 2011, 22(1):131-44.

[10] Mantas Luko?evi?ius. A Practical Guide to Applying Echo State Networks[M]// Neural Networks: Tricks of the Trade. Springer Berlin Heidelberg, 2012:659-686.

[11] Hermans M, Schrauwen B. Recurrent kernel machines: computing with infinite echo state networks.[J]. Neural Computation, 2012, 24(1):104-133.

[12] Rodan A, Tino P. Minimum complexity echo state network.[J]. Neural Networks IEEE Transactions on, 2011, 22(1):131-144.

[13] Zhang B, Miller D J, Wang Y. Nonlinear system modeling with random matrices: echo state networks revisited.[J]. IEEE Transactions on Neural Networks & Learning Systems, 2012,23(1):175-182.

[14] Alexandre, Luís A., Mark J. Embrechts, and Jamathan Linton. Benchmarking reservoir computing on time-independent classification tasks.Neural Networks, 2009. IJCNN 2009. International Joint Conference on. IEEE, 2009.

[15] Roseschies B, Igel C. Structure optimization of reservoir networks. Logic Journal of the IGPL,2010, 18(5): 635-669.

[16] Mitkowski P, Mitkowski W. Ergodic theory approach to chaos: Remarks and computational aspects[J]. International Journal of Applied Mathematics and Computer Science, 2012,22(2):259-267.

[17] Garfias P, Gutiérrez L, Sallent S. Enhanced DBA to Provide QoS to Coexistent EPON and 10G-EPON Networks[J]. Optical Communications & Networking IEEE/OSA Journal of, 2012,4(12):978-988.

[18] Sue C C, Chuang K C, Wu Y T, et al. Active intra-ONU scheduling with proportional guaranteed bandwidth in long-reach EPONs[J].Photonic Network Communications, 2014,27(3):106-118.

[19] Jun J H, Yu Z, Jung M S, et al. Inter-ONU Scheduling Scheme for QoS Guarantee in 10G EPON[C]// 2010 International Conference on Broadband, Wireless Computing, Communication and ApplicationsIEEE, 2010:137-141.

[20] Park K S, Yu Z B, Cho J H, et al. An Intra-ONU Scheduling Method in 10G EPON Supporting IEEE 802.1 AVB[M]// IEEE, 2010:631-636.