`ashift`

recommended = 12
vdev-scoped
- Immutable
Physical (on-disk) sector size
- ZFS stores data in records
  - Composed of sectors
Set in bits
- ashift=9 = 512B sectors
- ashift=12 = 4K sectors
- ashift=13 = 8K sectors
  - Some newer SSDs
If you get it wrong, overshoot - don’t undershoot

`recordsize`

dataset-scoped
- Mutable
Defaults to 128K (as of 2019)
- Pretty reasonable
Want to map close to workload size in dataset
- JPGs ~5MB
  - recordsize=1M
  - Prevents undue fragmentation
    - Fewer IOPS
- MySQL InnoDB
  - recordsize=16K
  - (InnoDB defaults to 16KB page size)
- MySQL InnoDB in a VM
  - KVM .qcow2 files default to a cluster_size of 64KB
  - recordsize=64K
Too High
- Increased latency
Too Low
- Greater fragmentation
  - Performance issues down the road
- Limit compression ability
- ZFS compression is by record
- compression=lz4, ashift=12, and recordsize=4K
  - No compression
    - Sector size is equal to record size
    - Can’t store data in half a sector
  - Default recordsize=128K
    - Could easily do 1.7:1
Torrenting
- Torrenting access is small chunks with random access
  - Suggests smaller record size
- Actual access is as larger files
  - Suggests recordsize=1M

xattr=sa
Sets Linux Extended Attributes directly in inodes instead of as files in hidden folders
Performance improvement on datasets with lots of files
- Especially with SELinux
Little effect on datasets with few large files