Describe the feature would like to see added to OpenZFS
The kernel ML project is using an in-kernel model to train on IO patterns and optimize access/code paths to suit actual needs. Extending this function or reproducing in a license-compatible manner into ZFS should provide similar benefits for DMU/ARC/etc, potentially coordinated with the underlying (block) IO layer (presuming object-style "vdevs" get in the mix before this sees light of day).
How will this feature improve OpenZFS?
Optimization of prefetch, ARC, block/IO sizing/batching/timing based on current workload (and possibly "profiles" like "backup activity")
Describe the feature would like to see added to OpenZFS
The kernel ML project is using an in-kernel model to train on IO patterns and optimize access/code paths to suit actual needs. Extending this function or reproducing in a license-compatible manner into ZFS should provide similar benefits for DMU/ARC/etc, potentially coordinated with the underlying (block) IO layer (presuming object-style "vdevs" get in the mix before this sees light of day).
How will this feature improve OpenZFS?
Optimization of prefetch, ARC, block/IO sizing/batching/timing based on current workload (and possibly "profiles" like "backup activity")