Sure you'll introduce a bit of Disk IO, but it's the same DiskIO for all of your test threads, so it should wash out in terms of a comparison.

Are you sure? The example data is far less than the size of a buffer, so won't Perl fill up a buffer and only then write it? That pattern will be predictable but it won't be uniform across those files.

Benchmarking is difficult; IO doubly so.