real    5m35.658s
user    98m5.770s
sys     5m58.114s

(31 workers)

I switched from line-oriented to block I/O, bumping the line count to 150 +
~30 (mainly for the higher-order function used to apply a function in a
different process). At this point, the speed is largely determined by the
residual data in the buffer cache, and there's a considerable variance across
executions.

Small improvements should be possible at the cost of higher line counts; there
are two obvious changes that might make this particular program faster:
single-copy read() (currently, the data is first read into a buffer on the
stack and then copied to the destination string) and mmap-based I/O.