I imagine we could get reasonable parallelism by limiting it to, say, 100 concurrent calls