This work explores the benefits of cloud computing in the development of kriging-based parallel optimization algorithms dedicated to expensive-to-evaluate functions. We first show how the application of a multi-point expected improvement criterion allows to gain insights into the problem of shape optimization in a turbulent fluid flow, which arises in the automobile industry. Our work then proceeds with a variety of experiments conducted on the ProActive PACA Grid cloud. Due to a multiplicative increase in search space dimensionality, the multi-point criterion cannot exploit a large number of computing nodes. Therefore, we employ the criterion with an asynchronous access to the simulation resources, when the available nodes are immediately updated while accounting for the remaining running simulations. Comparisons are made with domain decomposition which is applied here as an alternative parallelization technique. Our experiments indicate weaknesses in the use of the multi-point criterion with a synchronous node access, and benefits when working in the asynchronous mode. Finally, a relatively fast and accurate method is developed for the estimation of the expected improvement at multiple points.
