when a job has many input paths - listStatus - or the improved listLocatedStatus - calls (invoked from the getSplits() method) can take a long time. Most of the time is spent waiting for the previous call to complete and then dispatching the next call.
This can be greatly speeded up by dispatching multiple calls at once (via executors). If the same filesystem client is used - then the calls are much better pipelined (since calls are serialized) and don't impose extra burden on the namenode while at the same time greatly reducing the latency to the client. In a simple test on non-peak hours, this resulted in the getSplits() time reducing from about 3s to about 0.5s.