Abstract: The sample point generation strategy used in Latin Hypercube Sampling Monte-Carlo method is improved in according with the large variability of a concrete shrinkage and creep model. The improved strategy is developed under an unbiased mean value restriction; it can avoid negative values of random variables, so as to conform to the physical meaning. A simple example of shrinkage strain calculation shows that the improved strategy has an identical mean value and a slightly decreased standard deviation comparing to unimproved one. The probabilistic analysis is performed for a super long prestressed concrete frame structure; the changes of side column displacements and slab stresses in 10000 days are studied. The results prove that the improved sampling strategy can be applied successfully in the long-term performance probabilistic analysis for large and complex structures.