There is a validation paper on cluster-based correction methods that indicate that 800 iterations are enough to converge. So the default of 1000 iterations in CONN is fine, but of course for publication the more the better, so 10000 would be more than safe I guess.

Here is the paper: Pernet, C. R., Latinus, M., Nichols, T. E., & Rousselet, G. A. (2015). Cluster-based computational methods for mass univariate analyses of event-related brain potentials/fields: A simulation study. Journal of neuroscience methods, 250, 85-93.