Hi Bin Liu,
The threshold depends on the degrees of freedom. Say you want a cluster forming threshold that corresponds to a p = 0.001, and that your F-test has degrees of freedom of the numerator (i.e., of the model, or the rank of the contrast) equal to 2, and that the degrees of the freedom of the denominator (i.e., of the residuals, which generally is the number of subjects minus the number of EVs) is 50. Then in Octave (free) or Matlab (with the Statistics Toolbox) you obtain:
octave:1> finv(1-0.001,2,50)
ans = 7.9564
So, you'd use 7.96 as the cluster forming threshold.
All the best,
Anderson