Huh, you're right. When i make process_dump_file return a reference to the array and push that into an array outside, the benchmark suddenly looks like this:

Extracted 308272 orders.
CSV time: 3 wallclock secs ( 3.11 usr + 0.48 sys = 3.59 CPU)
Extracted 301468 orders.
CSV time: 7 wallclock secs ( 6.41 usr + 0.23 sys = 6.64 CPU)
Extracted 316912 orders.
CSV time: 7 wallclock secs ( 6.02 usr + 0.22 sys = 6.23 CPU)
Extracted 426854 orders.
CSV time: 8 wallclock secs ( 8.08 usr + 0.31 sys = 8.39 CPU)

Duration: 37 wallclock secs (32.28 usr + 2.36 sys = 34.64 CPU)
[download]

This is a bit weird, but already a lot better. Thanks for the suggestion. :)