In order to avoid cache thrashing, lfs migrate currently uses direct i/o to copy file contents, which may leads to significantly different performance compared to normal (non-O_DIRECT) i/o, cf. for example the discussion at http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/2017-November/015112.html. Internally, we've measured even more extreme differences where O_DIRECT throughput performed at 1/10th the normal rate. Therefore, I'd suggest to leave the choice between normal and direct i/o to the caller of lfs migrate by adding a new option N/-non-direct that suppresses to O_DIRECT flag in the open() call of the source file. Patch to follow.