Reference no: EM133779553
PROJECT PROPOSAL
Article - Load Balancing in Software-Defined Data Centre With Fat Tree Architecture
Implementation in Mininet- load balancing in SDN with Fat Tree Architecture using hashing, Markov and Q-learning(Reinforcement learning). All 3 are must(hashing, markov, Q-learning(Reinforcement learning).
Brief Summary of the Paper's Problem Domain/Challenge, Goals, and Technical Approach:
Paper have tried to resolve issue to balance load in efficient manner in SDDC by use of Fat Tree Network Topology. This technique is Cost effective as well as highly scalable. Hence finds wide use for the purpose. Challenge is to distribute the network traffic in an effective manner so that problem of Bottleneck does not arise , latency is reduced as well as high throughput is ascertained using Equi Cost Multi Path ( ECMP). Latest application such as Cloud computing need huge bandwidth which makes it a critical issue to balance the load in efficient way.
Paper have suggested various techniques of load balancing. Technique such as Hashing, Markov intend to achieve targets of Optimum performance in respect of Throughput , Latency as well as Network Resilience. Purpose is of Uniform Traffic Distribution.
Summary of the Paper's Current Implementation, Evaluation Strategy, and Key Results:
Paper have an idea about different strategies available in Fat Tree Topology to balance the load. Hashing technique is traditionally used to distribute traffic with consistent mapping of IP address of source and destination to decide as to through which server route the traffic. The method is simple and static but adaptability is limited. As such paper have explored scope of better technique such as Markov Decision Process.
This algorithm is more dynamic and more smart. This process enables Congestion aware routing as well as adaptive load balancing techniques. Results reveal that whereas use of static hashing is useful use of dynamic process as Markov is better as distribution of load is improved , congestion is reduced and overall network performance is improved.
Plan
Proposed Implementation (Language, Framework, etc.):
We have experimentally established correctness of processes suggested by paper in order to verify them. Because we have implemented the approaches suggested by paper.
Languages and framework used in implementation - Language - Python
Framework - Mininet for network simulation
Load balancing algorithm - Hash and Markov based approaches.
Environment - Fat Tree Topology using Mininet. Different Load Balancer shall be applied for management of traffic distribution.
Evaluation Strategy (Testing Platform/Setup, Simulated Data/Traces, etc.):
We will do evaluation by simulation of traffic within our Mininet Fat Tree Topology. We will generate traffic using Iperf and Ping for measurement of Throughput and Latency respectively. To observe performance of different Load Balancer under different Load conditions simulation of varied traffic loads such as Low , Medium and High has been done by our setup. Features of evaluation of Latency , Throughput and Packet loss under different Load conditions have been incorporated in our Key Evaluation Metrics.
Key Results Trying to Reproduce:
Instead of justifying results of approaches paper has suggested we have established correctness of results experimentally for both the techniques - Harsh and Markov.
Discussion
How You Can Compare Your Findings (Quantitatively, Qualitatively) with Previously Published Results:
Our Key Performance Metrics will provide quantitative comparison based on parameters of Latency, Throughput and Packet Loss. Metrics will also be compared for Hash versus Markov algorithms. Performance of different Load Balancer under different Load conditions will be plotted for quick observation and comparison.
New Questions/Settings Trying to Evaluate that Are Not Addressed in the Original Paper:
Future Scope - hope we will get time to reinforce our knowledge to be able to improve performance on parameters of Latency, Throughput and Packet Loss.
Evaluation of new question - effectiveness of Q-learning (Reinforcement learning) versus Hash/ Markov based load balancing. In paper ML based approach has not been used. ML based approach(Q-learning (Reinforcement learning)) can provide superior adaptability. Further original paper is emphasising on Congestion aware routing whereas our experiment will assess performance of Adaptive learning algorithms in varied load conditions. It will let us optimize continuously with time.