Since it appears to be a per-process limit, the way I have got over the limit in the past is to split the workload over a number of other processes. For example, with 128 jobs to run, spawn 4 programs which just run 32 jobs each.

Not elegant, I know.