Break large GET requests into a sequence of smaller ones about marfs

[jti-lanl]

[Old comment I've been meaning to post:]

You can sort of fake this in pftool, using e.g. "-C 1 -S 128M", to get it to assign sub-ranges. But those subranges will potentially get assigned to workers on different hosts, so they are requesting parts of the same object from different servers, which doesn't seem to work very well.

But we can fix that by changing the host-randomization to pick the server IP as a hash of the MD inode. Now, workers on different servers ask for subranges of the same object from the same server. But, those subranges still seem to be in an arbitrary order. So, that doesn't work very well, either.

[jti-lanl]

This is done in marfs 82182f. Not sure it if it is still useful, given our recent discoveries about network tuning, but one can turn it off by setting the new repo.max_request_size parameter to 0 (or by not providing it).