The latest technologies of server virtualization allow multiple computer workloads to share the same physical system dynamically while protecting them from interference from each other. Consolidation of workloads from stand-alone systems into virtualized systems requires accurate capacity sizing and optimal portfolio design to maximize the benefit of virtualization. This requirement often translates into a demand for accurate estimation of high quantiles from a limited amount of data for hundreds of workloads with diverse statistical characteristics. To deal with the problem, a semiparametric method of quantile function estimation is considered. The method employs the generalized Pareto distribution to model the high quantiles that exceed a certain threshold and retains the sample quantiles below the threshold. An automatic procedure is proposed for adaptive threshold and estimator selection and for adaptive data trimming. A simulation study shows that the procedure proposed is superior to the non-parametric sample quantile method for a variety of distributions. The procedure is applied to a portfolio optimization problem for computer workload consolidation with real data. © 2011 Royal Statistical Society.