On Mon, Jun 03, 2024 at 10:09:10PM -0700, Christoph Hellwig wrote: > On Mon, Jun 03, 2024 at 01:12:05PM -0700, Darrick J. Wong wrote: > > This patchset adds a different computation for AG count and log size > > that is based entirely on a desired level of concurrency. If we detect > > storage that is non-rotational (or the sysadmin provides a CLI option), > > then we will try to match the AG count to the CPU count to minimize AGF > > contention and make the log large enough to minimize grant head > > contention. > > Do you have any performance numbers for this? > > Because SSDs still have a limited number of write streams, and doing > more parallel writes just increases the work that the 'blender' has > to do for them. Typical number of internal write streams for not > crazy expensive SSDs would be at most 8. Not much other than the AG[IF] and log grant heads becoming less hot. That pushes the bottlenecks to the storage device, which indeed is about 8 per device. More if you can raid0 them. *Fortunately* for metadata workloads the logging code is decent about deduplicating repeated updates, so unless you're doing something truly nasty like synchronous direct writes to a directory tree with parent pointers that is being modified heavily, it takes some effort to overload the ssd. (Or a crappy ssd, I guess. Maybe I'll pull out the 860 QVO and see how it does.) --D
