Towards Approximating the Mean Time to Failure in Vehicular Clouds

Abstract

In a recent paper, Ghazizadeh et al. have studied vehicular clouds running on top of the vehicles in the parking lot of a major airport. The defining difference between vehicular clouds and their conventional counterparts is the unpredictable availability of computational resources. Indeed, as vehicles enter the parking lot, fresh compute resources become available; when vehicles depart, their compute resources leave with them. In such a volatile environment, the task of promoting reliability becomes quite challenging. To solve the reliability problem, Ghazizadeh et al. suggested employing redundancy-based job assignment strategies. They derived analytical expressions for the mean time to failure of these strategies. Their expressions require full knowledge of the distribution of vehicle residency times and of the time it takes to recruit a vehicle into the vehicular cloud. In a practical context, the datacenter manager does not know these distribution functions. Instead, using accumulated empirical evidence, she may know the first and perhaps the second moment of these random variables. With this in mind, this paper derives easy-to-compute approximations of the mean time to failure of the job assignment strategies proposed by Ghazizadeh et al. . A comprehensive set of simulations have shown that our approximations are very close to the analytical predictions by Ghazizadeh et al. even if the exact distribution functions are not known.