Cognitive Technique for Software Defined Optical Network (SDON)
1. Cognitive Technique for Software
Defined Optical Network (SDON)
Mônica de Lacerda Rocha
Electrical and Computing Engineering Department (SEL)
Engineering School of São Carlos (EESC)
University of São Paulo (USP)
monica.rocha@usp.br
18-19 May 2016
2. • Software defined optical network, SDON
• Concept overview
• Correlated research in our laboratory
• Cognitive optical network
• Concept overview
• Our Proposal
• Cognitive algorithm
• Case study
• Results and discussion
• Conclusion and future works
Outline
3. • Adaptive optical network
• Flexible optical transponder controlled by software
• Better allocation of resources in time (t) and
frequency (l) for different applications and
requirements.
• Higher spectral efficiency
• Thinner granularity for the connections
• Reconfigurable optical add-drop multiplexer
• Remote and dynamic traffic control
– Colorless
– Directionless
– Contention less
Software Defined Optical Network (SDON)
4. • Adaptive optical network
• Flex Grid
• Spectral optimization for capacity maximization
• Granularity of 12.5 (6.25) GHz to aggregate multiple optical subcarriers
[12.5 (6,25) GHz] in optical superchannels
• Self-adaptive optical amplifiers with gain adjustment for dynamic
flex grid operation
Software Defined Optical Network (SDON)
7. • SDONs require an intelligent control plane, capable of determining the
best route, best modulation format, best spectral grid, best FEC
scheme, etc, for each lightpath, which is accomplished by
Artificial Intelligence (AI) with learning capability
• This new approach, called cognitive optical network, aims to add
intelligence and bring autonomic operation to optical networks besides
providing advantages, in comparison with non-cognitive networks, such
as minimization of blocking probability, fastness to estimate the QoT, and
multiobjective optimization of parameters.
• SDON is the paradigm around which most of the research activities
conducted at the Optical Superchannel Laboratory (SEL-EESC-USP)*
are based upon.
Software Defined Optical Network (SDON)
*Electrical and Computing Engineering Department of the Engineering School of São Carlos (University of São Paulo)
8. • The laboratory is part of the Department Telecommunication Group
• Research activities are based on the optical superchannel concept,
that enable the deployment and operation of next-generation of
systems and networks.
• Our main challenge is to combine the research lines in an integrated
perspective where a data plane meets the requirements of
heterogeneous optical networks defined by software in a control plane.
• The research lines are mutually independent but are correlated and
may be aligned to, at least, one of three axes:
1. Data Plane
2. Control Plane
3. Planning
Optical Superchannel laboratory
http://www1.sel.eesc.usp.br/supercanal/eng
http://www1.sel.eesc.usp.br/supercanal/por
9. • Drs. Amílcar C. César (1) and Mônica L. Rocha (2)
• Drs. Daniel M. Pataca (3), Miquel A. Garrich (4) and Tania R. Tronco (5)
• PhD students: André L. F. Lourenço (6), Arturo M. Vera (7),
Natalia S. B. Capellari (8) and Rafael J. L. Ferreira (9)
• MSc students: Diego M. Dourado (10) and Leonardo A. Vanzella (11)
Research lines (correlated to this talk)
1 (USP) 2 (USP)
3 (CPqD) 4 (CPqD) 5 (FT-Unicamp)
6 (USP) 7 (USP) 8 (USP) 9 (USP)
10 (USP) 11 (USP)
10. Control Plane
• Shared Path Protection (SPP) Algorithm
• Goal: traffic protection and restoration in an elastic optical network (EON).
• The algorithm searches for primary and secondary disjoint paths.
• It divides the spectrum into two partitions and prioritizes slots in one of them
with secondary path traffic.
• It improves the blocking probability of connection requests, spectrum
utilization ratio, and average size of slot groups.
11. • Algorithm for traffic grooming in elastic optical networks
• The algorithm ZWNE (zone based with neighbor expansion) proposed for
WDM networks [Lee] is extended for optical elastic networks.
• An auxiliary graph is constructed for each connection request.
• The auxiliary graph initiates with a small region of the graph including a
candidate path between source and destination nodes.
• The region is continuously expanded until a path that satisfies the
requirements of the connection request is found.
• The technique reduces the computational time and improves engineering
traffic results.
Control Plane
[Lee] Q.-D. Ho and M.-S. Lee, “A zone-based approach for scalable dynamic traffic grooming in large
WDM mesh networks”, Journal of Lightwave Technology, vol. 25, no. 1, pp. 261–270, Jan 2007
12. • Algorithm for traffic protection and restauration
• The goal is to balance the choice between the position in the spectrum and
the chosen route in order to encourage spectral sharing in protection paths.
• This technique is faster than the scan with spectrum window.
• It chooses the solution with lower final cost.
• The algorithm scans the spectrum in all links of the route, slot by slot,
searching for a free band large enough to meet the demand.
• The logical search deals, simultaneously, with all the links in the route, by
performing logical operations with them.
• A convolution is performed between a spectrum window, with the same size
of the demand, and the resulting spectrum.
Control Plane
13. • All optical node architecture for optical OFDM operation
• Proposal and demonstration of an all optical Fast Fourier Transform (OFFT)
module for selecting any subcarrier of an optical OFDM superchannel.
• Proposal and demonstration of a node architecture for synchronous add-
drop multiplexing of subcarriers.
• Proposal and demonstration of an elastic optical network operation based
on optical OFDM.
Data Plane
14. • Planning strategies for increasing spectral and energy efficiency in
TWDM-PON and OFDM-PON
• The goal is to establish a compromise between energy consumption and
maximum bandwidth capacity.
• Scenario: a large bandwidth demand should be attended at the lowest
possible energy expenditure, without compromising the quality of service.
• Four classes of PON :
• GPON and XGPON, reference for performance comparison,
• TWDM-PON and OFDM-PON, where the algorithm is applicable.
• The algorithm can scale the distribution of users by optimizing the cost and
the quality of service (QoS).
• Results are promising for planning of sustainable access optical networks.
Planning
15. • Optical pulse shaping control in Nyquist-WDM systems
• Aiming the operation of flexible transponders, we study the impact of varying
some parameters in the transmitter/receiver module, such as
• Roll off factor
• Finite impulse response length of root-raised-cosine pulse shape
• Jitter
• DAC/ADC resolution
• The goal is to optimize the system performance by establishing a tradeoff
between impairments such as inter-symbol interference and crosstalk.
Data Plane
16. • Cognitive algorithm using fuzzy reasoning for software defined
optical network
• Proposal of a cognitive algorithm based on Fuzzy C-Means (FCM) technique
for the learning and decision-making functionalities of software-defined
optical networks.
• When included in a SDON control plane, the network achieves better performance,
when compared with a non-cognitive control plane
• As a case, FCM is applied for determining, in real time and autonomously, the
modulation format of high-speed flexible rate transponders in accordance with a
QoT standard.
• When compared to the case-based reasoning (CBR) algorithm, commonly used in
optical cognitive networks, FCM outperforms both fastness and error avoidance,
achieving 100% of successful classifications, being two orders of magnitude
faster.
Control Plane
Tania Regina Tronco, Miquel Garrich, Amílcar Careli César, and Mônica de Lacerda Rocha,
“Cognitive algorithm using fuzzy-reasoning for software-defined optical network”, Photonic
Network Communications, DOI 10.1007/211107-016-0628-1, published online: 16 April 2016
http://link.springer.com/article/10.1007%2Fs11107-016-0628-1
17. • A cognitive network comprises control mechanisms that may operate
in five steps:
1. observe and collect the information about the operation environment;
2. orient to evaluate the importance of the collected information;
3. learn from the experiences;
4. decide about which parameters/resources need to be (re)configured; and
5. act to adjust its parameters/resources.
• Steps (2) and (4) must follow end-to-end goals given by the network
operators such as performance improvements.
Optical Cognitive Network
18. • Our Problem: to choose an AI technique with capabilities for learning
with fast processing time and high precision in decision-making
Optical Cognitive Network
AI
Technique
Application Advantage Disadvantage
Case-Based
Reasoning, CBR
• Estimation of channel in cognitive
radio
• Estimation of QoT in optical
network
• Efficient spectral allocation in
wireless network
• Simplicity and
similarity to human
reasoning
• Learning based on
past cases
• Large data base
• Slow processing time
• Does not solve
multiobjective problems
Artificial Neural
Network
• Spectral prediction and channel
selection in cognitive radio
• Adjustment of optimum operation
point in a cascade of optical
amplifiers
• Low use of memory
• Fast response
• Excellent for pattern
classification
• Requires training
• Output is not trackable
• Complex processing for
training
Genetic
Algorithm, GA
• Wireless network optimization
• Routing with restrictions of QoS
• Dynamic optical networks
• Protection and restoration in
optical
• Parallel processing
• Requires little
knowledge of the
technology
• Slow processing time
19. • Fuzzy C-Means, FCM, successfully applied in cognitive radio, could
fulfil the requirements by being able of
• Learning
• Automatically generating rules, from data provided by monitors (spread in
the network) and simulators
• Dynamically changing the rules, as new data are aggregated to the system
• Fast and precise decision-making
• Based on that, we have proposed, for the first time in optical
networking, the use of FCM
• We then studied a case for determining, in real time, the modulation
format of flexible transponders, and compared the FCM performance
with CBR
• Finally, we propose a new control plane architecture that includes FCM
and a more complete definition for optical cognitive network, in this
context
Optical Cognitive Network
20. • FCM is an hybrid algorithm, resulting from the combination of fuzzy
logic (fuzzy) with the data clustering method (C-Means).
• FCM was proposed in 1981 by Bezdek.
• It has been used for pattern recognition and, more recently, with
effective gain in respect to the CBR algorithm, for cognitive radio*
aiming
• Radio channel estimation
• Spectrum allocation
• Modulation format
Fuzzy C-Means
* H. Shatila, “Adaptive radio resource management in cognitive radio communications using fuzzy reasoning”, Ph.D. dissertation, Virginia
Politechnic Inst. and State Univ., 2012.
21. • The purpose of the data clustering method (clustering) is:
• To group similar data set in different clusters
• To identify such clusters in unsupervised mode
• Unsupervised mode: no information is provided, in advance, to the algorithm
about which data belong to which groups;
Data Clustering Method
x
x
x
x
xx
x
x
x
22. • In Boolean Logic, an element belongs or does not belong to a set.
• In a Fuzzy logic algorithm, the knowledge is represented by means of
IF…THEN rules.
Fuzzy set theory differs from traditional set theory, where either an element
belongs to a set or it does not.
In FL, a partial membership is allowed, i.e., an element can belong to a set
only to a certain degree.
This membership degree is usually referred to as the membership value and
is represented by a real value in the interval [0, 1], where 0 and 1 correspond
to full non-membership and membership, respectively.
Fuzzy Logic
23. • Fuzzy clustering system for classification includes the following steps:
(1) collect data from the system, by measuring or via computer simulations;
(2) determine the model structure suitable to the problem by identifying the
relevant characteristics and selecting, from the collected data, the proper data
for training the algorithm;
(3) select the number of the required clusters;
(4) cluster the training data using FCM algorithm;
(5) obtain the membership functions from the clusters;
(6) determine the fuzzy rules from each cluster by using the obtained
membership functions; and
(7) use the fuzzy rules to configure the system.
• FCM membership functions are estimated from stored training data, and hence, the
cognitive engine is learning from experience.
FCM Algorithm
24. • We applied FCM to determine the modulation format to be used
according to a given QoT, for a connection request of 200 Gb/s, that
may be provided by varying the number of subcarriers and the
modulation format of an optical OFDM stream of data.
• DP-16QAM 200 Gb/s
• DP-QPSK 100 Gb/s
Case study
Simulation flow chart
(modulation format
determination using
CBR and FCM schemes)
25. • The training data are obtained from previous off-line computer
simulations performed using the OptiSystem simulator 13.1 and
considering optical transmitters and receivers (setup with the two
modulation schemes, i.e., DP-16 QAM and DP-QPSK), a coherent
receiver, a digital signal processing module, Erbium-doped fiber
amplifier, a Gaussian optical filter, electrical amplifiers, and a standard
optical fiber.
Link Simulation
26. Training data
Note: the same training data are used to build a KB (knowledge base) for a
CBR algorithm.
27. Training data
The figure illustrates the belongingness to
a cluster as a function of
(a) Route length (input)
(b) modulation format (output)
Rules
1. If the route is in cluster 1 then the
modulation format is in cluster 1
2. If the route is in cluster 2 then the
modulation format is in cluster 2
28. Optical Network
• For the computational simulations of an optical network a generic
long-haul eight-node mesh topology.
• The performance of the FCM and CBR algorithms was compared in
terms of computational time and accuracy to take decisions about the
proper modulation format to set up.
30. Results and discussion
• Performance comparison between
FCM and CBR - Computational time as
a function of number of connections
• FCM after 2500 connections:
average computational time: 14.4 seconds,
(standard deviation of 0.2)
lower and upper limits of 95%
(confidence interval of average:14.2 and
14.5, respectively).
• CBR after 2500 connections:
average computational time: 1405.6sec
(standard deviation of 33.1)
lower and upper limits of the 95%
(confidence interval of average: 1381.9 and
1429.2, respectively)
31. Results and discussion
• Performance comparison between FCM and CBR - Computational time
as a function of number of connections
FCM is around two orders of magnitude faster than the CBR when 100
training cases are used.
Both algorithms provide 100% of agreement in the modulation format
selection.
If the number of training cases is reduced to 50, the FCM continues to provide
100% of successful selections, while CBR presents an error percentage
around 30% for 5000 connection requests for both methodologies.
• These errors occur due to the reduction in the number of cases in the KB
(low granularity).
32. Results and discussion
• Performance comparison between FCM and CBR - Computational time
as a function of number of connections
• The mean computing time to select the modulation format with FCM is
6.47 ms, (same order of magnitude obtained by Jimenez et al.* with
operation in real time).
• FCM does not require a database to store known cases and wasting time
to search similar cases in this database, neither needing to use learning
and forgetting techniques to optimize this database.
• FCM allows the inclusion of other co-related parameters, with relative
simplicity, by just including new FCM rules.
Jiménez, T., et al.: A cognitive quality of transmission estimatorfor core optical networks. J.
Lightwave Technol. 31(6), 942–951 (2013)
33. Results and discussion
• Relationship between computational
time of FCM and CBR
We set the number of connection requests to
500 and change the number of cases stored,
N, from 100 to 50,000, to compute the
relationship between computational.
These results are in agreement with the
predicted in the time complexity analysis
described in [1], which provided
a linear time complexity for CBR and a
constant one for the FCM algorithm.
That proves that the FCM technique
is faster than CBR and the number of stored
cases directly influences the performance of
the CBR.
[1] Tania Regina Tronco, Miquel Garrich, Amílcar Careli César, and Mônica de Lacerda Rocha, “Cognitive algorithmusing fuzzy-reasoning
for software-defined optical network”, Photonic Network Communications, DOI 10.1007/211107-016-0628-1, published online: 16 April
2016
35. New SDON Definition
A cognitive SDON is a software-defined optical network intelligent and aware
of its QoT, of its spectrum availability, of service requirements, and of energy
saving and security requirements, which follows policies given by network
operators.
It uses a learning technique to learn from cases in the past and adapt its
internal states (configurations) as a function of changes in the optical
medium, by adjusting, in real time and autonomously, its parameters of
operation—bit rate, modulation format, FEC scheme, wavelength, numbers
of frequency slices, add/drop channels, number of optical carriers—in order
to achieve a high-quality communication, high availability, and efficient
utilization of the optical spectrum.
36. Conclusion and future works
• New approach based on FCM that, as far as we know, has been
applied for the first time in a SDON context.
• Case studied (real time selection of modulation format to a
certain lightpath) with 100% of successful assessments
• FCM is much faster—close to two orders of magnitude—than a
traditional CBR algorithm and bringing additional advantages,
while maintaining good performance and scalability.
• We have focused on off-line training, but it is possible for the
algorithm to adapt itself, in real time, to a changing environment,
when working together with an OSNR monitoring system.
• That is feasible because the time processing to adapt the
membership functions with the new data collected by the monitor is
very low.
• Additionally, we proposed a definition for a cognitive optical
network and an architecture for the SDON control plane that
includes the FCM algorithm.
37. Conclusion and future works
Future works include:
(1)to carry out simulations of propagation through cascades of
ROADMs and optical amplifiers;
(2)to analyze the performance based on more flexibility on the
number of modulation formats, bit rates and subcarriers;
(3)to compare the performance of FCM technique to other artificial
intelligence techniques, such as neuro-fuzzy;
(4)to validate the FCM algorithm in a SDON control plane platform
using OpenFlow; and
(5)to develop a spectrum allocation algorithm based on FCM
technique..