Automatically Generated Simulations for Predicting Software-Defined Networking Performance
0 likes44 views
This document proposes a new modeling language called Network Modeling Language (NML) to model software-defined networking (SDN) and automatically generate simulations using different performance models like queueing Petri nets (QPN) and stochastic Petri nets (SPN). It presents the NML metamodel and model transformations from NML to QPN and SPN. Experiments comparing the accuracy and simulation time of the generated QPN and SPN models against real measurements on a SDN testbed show that QPN predictions have 3% error while SPN simulations are faster but less accurate. The work contributes an approach for automatic performance prediction of SDN networks from a single high-level model.
![9/30
Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions
Network Modeling Language
We propose the Network Modeling Language (NML) to create
SDN models and evaluate their performance. It is an extension to
our prior work on the Model-Driven Networking (MDN) framework
[1] and its metamodel:
Lopes, F. A. et al. CIn-UFPE | IFAL
Automatically Generated Simulations for Predicting Software-Defined Networking Performance](https://image.slidesharecdn.com/slides-210710160416/85/Automatically-Generated-Simulations-for-Predicting-Software-Defined-Networking-Performance-9-320.jpg)
![13/30
Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions
NML-to-QPN Transformation
From the formal definition present in [2], we designed
NML-to-QPN transformation as follows:
1 Each NML’s Host instance is transformed into a QPN’s place
p. The network interfaces of each Host are transformed into
a queueing place q;
2 Instances of NML’s Link and Switch are transformed into q
and transitions t. The Link capacity is represented as a weight
wi , which may also contain a color c to represent the delay;
3 Instances of Traffic and Flow define the set of colors C and
the number of tokens (according to the flow size)
4 The Controller and Rule entities are transformed into q and
immediate transitions W̃2.
Lopes, F. A. et al. CIn-UFPE | IFAL
Automatically Generated Simulations for Predicting Software-Defined Networking Performance](https://image.slidesharecdn.com/slides-210710160416/85/Automatically-Generated-Simulations-for-Predicting-Software-Defined-Networking-Performance-13-320.jpg)
Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 1. 1/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Automatically Generated Simulations for Predicting Software-Defined Networking Performance Felipe A. Lopes1,2, Rafael R. Souza1, Stenio Fernandes1 1Informatics Center, Federal University of Pernambuco, Recife, Brazil 2Federal Institute of Alagoas, Arapiraca, Brazil IEEE Symposium on Computers and Communications 25-28 June 2018 // Natal, Brazil Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 2. 2/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 3. 3/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 4. 4/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Context Several modeling and simulation approaches can be used for predicting network performance. However: Each approach has its own formalism and accuracy rate; Their use requires different expertise and manual steps. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 5. 5/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Motivation SDN network performance could be represented in multiple performance models, e.g., Petri nets (PN), analytical models, and queuing networks. Choosing the more accurate model is not a trivial task (accuracy, simulation time). Using models at design-time enable us to correct design errors before the deployment of the real network. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 6. 6/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Problem definition Research Question How can one obtain accurate simulation/prediction models for network performance without the need of different pieces of knowledge and experiences when designing a network? Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 7. 7/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Objectives Propose new model-to-model transformations that enable predicting performance metrics based on different performance models through a single high-level model instance. Evaluate stochastic models considering their accuracy in predicting an SDN network performance. Demonstrate the challenges and benefits of using different simulation models to represent SDN scenarios. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 8. 8/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 9. 9/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Network Modeling Language We propose the Network Modeling Language (NML) to create SDN models and evaluate their performance. It is an extension to our prior work on the Model-Driven Networking (MDN) framework [1] and its metamodel: Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 10. 10/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Meta-models Figure: Short version of the NML’s metamodel. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 11. 11/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 12. 12/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Model Transformation NML instances are transformed into specific PN models by using a Model-Driven Development (MDD) approach (i.e., Epsilon framework). So far: Queueing Petri Nets (QPN) and Stochastic Petri Nets (SPN). Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 13. 13/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions NML-to-QPN Transformation From the formal definition present in [2], we designed NML-to-QPN transformation as follows: 1 Each NML’s Host instance is transformed into a QPN’s place p. The network interfaces of each Host are transformed into a queueing place q; 2 Instances of NML’s Link and Switch are transformed into q and transitions t. The Link capacity is represented as a weight wi , which may also contain a color c to represent the delay; 3 Instances of Traffic and Flow define the set of colors C and the number of tokens (according to the flow size) 4 The Controller and Rule entities are transformed into q and immediate transitions W̃2. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 14. 14/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 15. 15/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Evaluation methodology 1. Experiments Model transfomation (i.e., NML-to-QPN, NML-to-SPN). Metrics verified: Accuracy of each model (e.g., throughput); Simulation time. 10Mb 10Mb Controller Host:h1 Switch:s3 dst: h2 size: 1000Mb protocol: TCP App:LoadBalancing algorithm: round-robin switches: {s2, s3} match: pkt.src == h1 condition: if (s2.out_speed < 6.5) Rule:A Flow:A Host:h2 Switch:s2 Switch:s1 Switch:s4 Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 16. 16/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Evaluation methodology 2. Testbed Our testbed was a commodity hardware with an Intel Core i7-5500U 2.40GHz processor and 8 GB of memory. Mininet. Ryu Controller. D-ITG. 3. Simulation We compared the prediction results of each model with a Mininet simulation. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 17. 17/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Results h1 h2 h1-port1 s1 s1-controller controller s1-output s2 s3 s2-s3-output s4 s4-output Queueing Place Ordinary Place Immediate Transition h1 h2 h1-port1 s1 s2 s3 s2-output s4 s4-output 7 7 7 7 7 7 s3-output s1-output-s2 s1-output-s3 rule1-1 rule1-2 Exponential Transition a b #rule1-1 > #rule1-2 #rule1-1 <= #rule1-2 Transformation Our transformation considers the conditional rules present at the NML model, mainly for the NML-to-SPN transformation. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 18. 18/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Results 0.0 0.2 0.4 0.6 0.8 1.0 5 6 7 8 9 10 Throughput (Mbps) Measured QPN Simulation SPN Simulation Figure: CDF for the throughput obtained (simulated and predicted). Accuracy In this comparison, our generated QPN model achieved an error rate of only 3%. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 19. 19/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Results Depending on the situation, a network designer or operator could need a less accurate but faster prediction. Simulation time QPN model (more accurate) took a mean of 5.16 seconds. SPN finished its simulation in 3.2 seconds (mean). Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 20. 20/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Outline 1 Introduction 2 Network Modeling Language 3 Model Transformation 4 Experiments and Results 5 Conclusions Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 21. 21/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Contributions A new model transformation to predict SDN performance. Identification of QPN as a more accurate prediction model (compared to SPN). SPN provides more realistic results when conditional rules are defined by an SDN application. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 22. 22/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Future work Integrate the prediction models into an autonomic architecture for SDN. Implement and test new transformations or prediction algorithms. Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance
- 23. 23/30 Outline Introduction Network Modeling Language Model Transformation Experiments and Results Conclusions Thank you! Questions? You can also contact me: fal3@cin.ufpe.br Lopes, F. A. et al. CIn-UFPE | IFAL Automatically Generated Simulations for Predicting Software-Defined Networking Performance