I think it is better to use reml to estimate the variance components and use these output estimates as starting values for programs based on bayesian analysis like (gibbsf90) and work with chain samples
I don't know if this is still available, but a long time ago I used MTDFREML (Multiple- Trait Derivative Free REML), a program written by van Vleck and van Tassel. Just checked - it appears to be still available at http://aipl.arsusda.gov/curtvt/mtdfreml.html