Service Abstractions With Fault Virtualization for Distributed Network Infrastructures
DS-RT '13 Proceedings of the 2013 IEEE/ACM 17th International Symposium on Distributed Simulation and Real Time Applications
| Hi-index | 0.00 |
Popular web sites receive an enormous share of Internet traffic. These sites have a competitive motivation to offer better service to their clients at lower cost. One of the solutions is to using content distribution network (CDN). When we optimize proxy placement in content distribution network we want to maximize the user experience in the mean time to minimize the resource consumption. We are dealing with a multi-objective optimization problem for CDNs in the thesis: the content latency experienced by users and the system resources expended to deliver content. After the topology model for content distribution network, we listed metrics to evaluate content distribution network. Then we defined the objective function for optimizing proxy placement. After reviewing the existing proxy placement algorithms we used genetic algorithm to solve the proxy placement problem. We first apply the genetic algorithms to a simple network, and find that the results are better than greedy algorithm and identical to the global optimal. Then we apply the genetic algorithm to a realistically larger network topology. The genetic algorithm can find good solutions to the proxy placement problem. But compared to greedy algorithm, it is less efficient. So when response time is important, genetic algorithm is not the best choice. To optimize the genetic algorithm we found if we choose higher mutation rate and step size at the beginning of the genetic operation and decrease the mutation rate and step size as we go through the genetic operation, the results are much better. For the initial populations, if we want quick response to the sudden change of client access we can choose smaller initial population, if we have sufficient response time especially in initial setting up the network we should choose larger initial populations. Under optimized genetic algorithm, we experiment three different network examples with their sizes between the simple network and large network. Since the network sizes are smaller we are able to run the exhaustive search to find the global optimal. In all cases the results from genetic algorithm outperform greed algorithm and close to global optimal. Although it is not guaranteed to find the global optimal solution as the simple network did, a robust good solution is enough to solve real world problems.