Mu haplotype caller is slow. Can I break the reference assembly into small files(20filtes, each contain several thousand contigs) and run the haplotype caller again?
Not explicitly -- have you looked at the presentations from our recent Queue workshop? http://www.broadinstitute.org/gatk/guide/events?id=3391
Yes, or you could use Queue to scatter-gather. Be aware that GATK has issues dealing with very large numbers of contigs. It is not designed to handle things like draft genomes, so performance will definitely be lower than for "finished" genomes.
I read the queue docs, but it's still confusing, are there some examples to run haplotype caller use Queue?