There are two options in TASSEL for variance component estimation while running a mixed linear model (MLM) for association mapping. First one is to use P3D compression, second is to re-estimate variance after each marker.
TASSEL help window claims that the compression saves time and overall model fit. Furthermore, it is said that for small and moderate-sized data sets any of the methods should give reasonable performance. If you are you are well versed in this topic please help to understand the idea and how to choose appropriate option.