Ran a couple more profiles. Looks like we're in pretty good shape. The vast majority of the time is being spent in the kernel doing file IO and messaging stuff, which is appropriate for this process. This process's raison d'etre is to do file IO and messaging stuff.

If this suggests anything, it's that if we spot an opportunity to do file IO smarter we should take it. Maybe a file descriptor cache. Maybe consolidate reads/writes if we see a place where we are artificially splitting them up.

That is a deeper investigation though, will have to do it later. I'm happy with the result for now.