In mobile clouds, users accessing cloud services on mobile devices may experience long delays that are detrimental to the user experience. One solution largely investigated is to deploy cloud services on cloudlets, smaller but powerful servers located close to the mobile users. While solving the latency problem, this new model raises two new challenging questions: how to achieve an optimal placement of services on cloudlets and how to minimize the placement transitions of services on cloudlets, both driven by the mobility of the users.
This paper addresses these two problems via a thorough analysis of the service placement problem in the context of a cloudlet-cloud architecture, followed by a presentation of original algorithms that aim to optimize the tradeoff between service latency, cloudlet resource usage, and service placement transitions. The first algorithm is a greedy one that determines the service placement and load dispatching in two steps. The first selects cloudlets as the destination for services, followed by load assignment. Three other algorithms are used as benchmarks.
The next contribution of the paper is an online solution that tries to solve the problems in real terms. This solution works with load prediction for each new period of time, considering previous load profiles. The optimization module decides on service placement and load dispatching. All solutions are clearly and rigorously presented.
The performance evaluation consists of two stages. The first set of algorithms is run considering a realistic set of simulation parameters. The greedy algorithm outperforms the others when the number of user requests varies from 1000 to 3500, and the number of services changes from 15 to 50: the latency by using cloudlet varies from 1/5 to 1/2 in respect to the latency when only the cloud is used. For testing the online solution, the authors considered a real-world dataset--Shenzhen taxi and metro data collected during one day in 2014. First, the accuracy of load prediction is considered, followed by the analysis of latency. The results are interesting although not impressive: latency can be up to 40 percent of the cloud access latency.
Summing up, this is an interesting paper that discusses a hot topic and proposes an initial set of solutions. The evaluation methodology is appropriate for the time being, but will require more real dataset experiments. This paper can be an eye opener for young researchers and also mobile service providers.