Instead of a permutation test, a Fisher exact test or hypergeometric test is more commonly used to calculate gene set enrichment.

When doing something like GO enrichment (which should use a similar principle), I don't set a hard cutoff for number of genes in the original gene set (in GO), but I typically like to see highly significant values (such a p<1e-5) that should typically include multiple enriched genes within the deferentially expressed gene list (similar to your 2000 gene list, I assume). However, I return the entire list of results p<0.05. Sometimes biologists like to know if a single gene is affected (if that single gene is known to be really important).

BTW, you can try using the TRANFAC enrichment tool in GATHER if you have a list of official gene symbols:

http://gather.genome.duke.edu/

I personally like the upstream regulator function in IPA (based upon literature annotations rather than predicted motifs), but that is commercial software.