On Fri, Jan 31, 2025 at 03:02:09PM +0530, Kundan Kumar wrote: > > IOWs, having too much parallelism in writeback for the underlying > > storage and/or filesystem can be far more harmful to system > > performance under load than having too little parallelism to drive > > the filesystem/hardware to it's maximum performance. > > With increasing speed of devices we would like to improve the performance of > buffered IO as well. This will help the applications(DB, AI/ML) using buffered > I/O. If more parallelism is causing side effect, we can reduce it using some > factor like: > 1) writeback context per NUMA node. > 2) Fixed number of writeback contexts, say min(10, numcpu). > 3) NUMCPU/N number of writeback contexts. Based on Dave's feedback, it would seem not using 4) can in the worst case make things worse in certain heavy workloads. So an opt-in rather than default would probably be best for 1-3. > 4) Writeback context based on FS geometry like per AG for XFS, as per your > suggestion. To this later point 4): This is not the first time having the ability to gather filesystem topology somehow comes up for more interesting enhancements, FDP being the other one. I don't think we have a generic way to gather this information today, and so do we want a way to at least allow internal users to query for something like this? Luis
