On Sun, Oct 20, 2024 at 02:30:43PM -0700, Linus Torvalds wrote: > On Sun, 20 Oct 2024 at 14:29, Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote: > > > > That's the same as limiting the amount of dirty metadata. > > Exactly. > > We limit dirty data for a reason. That was what I said, and that was my point. > > Excessive dirty data is a huge latency concern. It's a latency concern > not just for journal replay, it's a latency concern for anybody who > then does a "sync" or whatever. And my counterpoint is that if you've got a huge filesystem, limiting dirty metadata means that you force it to be written out as a whole bunch of random tiny writes, instead of allowing them to be batched up. That's horrible for overall throughput.
