I was about to complain a bit about the background here - eg. it’s true that some “modern storage” is very fast, specifically where “modern” means “latest SSD metal hardware”. But “modern” could also mean “cloud disks”; ie. to me it makes me think of GCPs yet-to-be-out-of-beta NVMe disks with battery-backed controllers.
And I was going to go on about how, for cloud disks, the latencies I’m seeing for page faults are more like 1ms than 1us.
But then I saw the link to https://www.datadoghq.com/blog/engineering/introducing-glommio/ and got super excited, that seems really cool; I’ve been keen to find a somewhat higher level library for io_uring on Rust. This may be too high level, but still a really cool looking project to explore!
This is pretty cool, really. Since a fair chunk of the performance improvement seems to come from OS paging increasing the amount of bookkeeping required, I wonder if telling your OS to use larger pages would also make any difference?