DirectIO For OpenZFS Shows Very Promising Performance

[jochendemuth]

Very neat. I did chuckle when I noticed they were benchmarking with a 90 disk array there at the end of the slides. I assume this will have next to no benefit for those of us with a spinny disk mirror or a single NVMe drive?

This is a most important note. Single disk configurations, general desktop/workstation use cases don't really benefit from this. ZFS works great already. So do all kinds of other file systems.

As I wrote earlier - I am exploring configurations using RDBMS (e.g. PostgreSQL) with multiple NVMEs for potentially highest perforamance, great performance/$ ratio setups. ZFSs snapshot capabilities are extremely valuable for online backups. ZFS+DirectIO should allow for improved storage performance using tuned synchronous_commit configurations.

I am very excited about this development and looking forward to the next ZFS version with this capability.

[coder]

any io_uring support on the roadmap ? I think it could increase perf too without need to use O_DIRECT.

It should already just work. That's the beauty of io_uring.

As for whether it's an effective substitute for using O_DIRECT, that really depends on the nature of the performance bottlenecks they hit and the sort of access patterns they're trying to optimize for. I can't say without seeing the presentation, which is more than I plan to invest in this subject.

My only reason for commenting here was to remind people of all the unhappy tradeoffs you're making, when using O_DIRECT. It's a big hammer and I frankly think the dichotomy is more drastic than necessary.

[deusexmachina]

https://github.com/openzfs/zfs/issues/8381 maybe zfs shouldn't be agnostic and should automatically apply optimizations on disk type. NVMes should be way faster on ZFS by default. So many use ZFS on rootfs of their workstation OS.