See "Inference from simulations and monitoring convergence1" by Gelman and Shirley. The especially point out that quite few McMC samples are needed to get a reasonable picture of posterior distribution, while more samples are needed if the precise knowledge of posterior mean or any similar quantity is needed.
