Zoned Namespaces

ZNS changes the storage interface, not the underlying cell technology. A zoned namespace is a logical address space divided into sequential-write-required zones. ZBC/ZAC host-managed disks may expose conventional random-write zones. ZNS does not define that zone type; a controller can expose a separate conventional namespace. [1]

In the ordinary sequential model, regular writes start at the zone write pointer. The host chooses when to retire a zone, while the SSD still handles media reliability. A zone is neither a raw NAND block nor a direct physical-address API. [1] [4]

Zone size is the address span; zone capacity is the writable block count. Capacity can be smaller, leaving an unwritable tail before the next zone. Count usable space from reported capacities, and check open and active resource budgets separately from available bytes. [1]

Where supported, Zone Append lets the controller choose a write location and return its starting LBA on successful completion. Concurrent appends need not be placed in submission order. Record the returned addresses and recover application mappings; a successful append does not establish a multi-request transaction. [2] [7]

Close preserves a partially written zone for further writes but retains an active slot. Finish makes it full, giving up unused writable capacity until reset. Reset returns an eligible zone to Empty and invalidates prior contents for normal access; this is not a secure-sanitization assurance. [2] [7]

An allocated Zone Random Write Area (ZRWA) permits non-sequential writes within a bounded window. Support and allocation must be checked explicitly. Under revision 1.5, an implicit ZRWA flush must occur even if its triggering write fails; an error is not proof of an unchanged pointer. The ordinary sequential model above assumes no allocated ZRWA. [3]

The ATC 2021 study adapted f2fs and RocksDB/ZenFS rather than merely changing the SSD label. Its mapping, spare-space and reclamation choices are part of the result. Database compaction can still rewrite data even when the device needs less internal relocation. [4]

The eZNS study found interference between zones on its tested commodity SSD, including shared write-cache effects. Treat that as evidence against assuming isolation, not a latency prediction for every device. Evaluate the actual firmware, workload mix and host allocation policy. [5]

A compatible filesystem can absorb much of the placement work. zonefs instead exposes zones as files and requires append-only direct writes to sequential-zone files. After partial I/O failure, reconcile reported state and validated records before retrying or reclaiming space. [6]