A comparative analysis of centralized and distributed load balancing in a software-defined networking environment
Date of Award
Master of Science in Computer Science
Information Systems & Computer Science
Yu, William Emmanuel S., Ph.D.
In a traditional hardware-level approach, implementing a distributed load balancer is extremely difficult due to cost, vendor dependency, and compatibility of each hardware component with the distributed network environment. Because of this, many enterprises are struggling to make a transition from centralized load balancing to distributed load balancing. With emerging technology of Software-defined networking, implementing a distributed load balancer is more cost-effective and easier to implement because components can be managed without relying to vendor-specific configurations. This study aims to develop a viable framework for distributed load balancing that does improve or does not negatively affect the overall performance of the load balancing process compare to the ones in centralized load balancing. Star topology and Clos Network Model were used as testbeds for centralized and distributed network environments and these were tested on a virtual machine. Six load balancing algorithms, namely, round-robin, IP hashing, weighted random, least connections, least packets, and least bandwidth, were integrated in both networks. Results showed that overall throughput drastically increased for distributed network on both TCP and UDP and all load balancing algorithms. The test was done on two virtual machines, namely, Amazon Web Services EC2 instance and Google Cloud Platform Compute Engine instance. For distributed network in TCP, a decrease of -40% to -70% in connection establishment delay, an increase of -200% in network throughput, an increase of -100% to -200% in total packet reads, and a decrease of -40% to -80% in total packet retries. For distributed network in UDP, a decrease of up to -75% in connection establishment delay, an increase of -200% in network throughput, a decrease of -45% to 70% in jitter, and a decrease up to -70% in packet loss. There was no single load balancing algorithm that clearly outperformed or underperformed other algorithms in terms of performance. The t-test results showed that the both networks had a p-value lower than the suggested value of 0.001. Future works like testing the framework in a physical environment and implementing multiple controllers in the network can be done in order to improve the inconsistencies that was made in this study.
(2018). A comparative analysis of centralized and distributed load balancing in a software-defined networking environment. Ateneo de Manila University.