A Comparative Study of Dynamic Resource Allocation Techniques in Cloud Computing: Load Balancing in Compute Power

Abstract

This thesis aimed to undertake a research experiment on two widely used cloud computing dynamic load balancing techniques, namely dynamic centralized load balancing and dynamic decentralized load balancing techniques, by comparing their performance in real-world applications. The thesis also looked at the single-switch topology as the baseline upon which most related POX load-balancing experiments are designed. The load-balancing algorithms utilized in the experiments included random, round-robin, weighted-round-robin, and least connections. The load-balancing performance comparisons were done using the Siege HTTP tool, which tracks, among other parameters, transaction rate, throughput, and response time. The load test variables passed into the tool were 10, 50,100,200 concurrent users/requests. The mean results demonstrated better performance indicators among the single-switch topology parameters than did the centralized and decentralized topology parameters combined. Then, in terms of the centralized and decentralized comparison, respectively, the former performed better than the latter in terms of average response time and average throughput except for the weighted round robin (WRR) algorithm and better than the latter in terms of the average transaction rate, except for round-robin (RR), and WRR algorithms, respectively. These results demonstrated the performance variability relative to the complexity of the software-defined networks, with the decentralized network paying the heaviest casualty for the most part among the parameters being analyzed due to the number of network elements present in the topology relative to the others except where additional intrinsic algorithmic logic such as WRR, and RR, for example, compensated for the network complexity. From these results, we concluded that the level of SDN topology complexity and the SDN Controller framework used, respectively, impacted the overall performance of the load-balancing algorithms, at least in terms of the parameters examined. Also, from past related studies and our study results, the POX controller framework seems optimized only for single-switch SDN networks. Future studies could, therefore, vary the loads even further, add performance tuning to all three topologies, add extra algorithms to the stable, expand more parameters for analysis, and switch to controllers other than POX, particularly for the centralized and decentralized frameworks