It depends on the length of the repeated parts of your genome. Repeated parts result in collapsed contigs (with subsequently higher read depths) that break the nonrepeated parts. Paired end reads where the distances between the halves are longer than the repeat size allow for ordering and orienting the nonrepeated contigs, and occasionally place copies of the repeats as well.
So, you want the longest paired end distance to be somewhat larger than the longest repeated parts (contigs). How to find this out? Look for the length of the highdepth, i.e. repeated, contigs, as written in the 454ContigGraph (first part, where all the contigs are listed). Very helpful is to plot the length versus depth, then you can match the paired end distances with the length of the repeated contigs.
Hope this helps, and hope you find out you don't need 8 kb or 20 kb...
flxlex
