Venue | Category |
---|---|
HotOS'21 | Zoned Storage |
Don't Be a Blockhead: Zoned Namespaces Make Work on Conventional SSDs Obsolete1. SummaryMotivation of this paperZNS SSD 2. Strength (Contributions of the paper)3. Weakness (Limitations of the paper)4. Some Insights (Future work)
Motivation
Industry has standardized and is adopting Zoned Namespaces (ZNS) SSDs, which offer a new storage interface that dominates conventional SSDs
ZNS devices dominate
Zoned namespace SSDs in six different states: empty, open, closed, full, ready-only, and offline
ZNS costs less per GiB
Less on-board DRAM: the FTL maintains address translations at the granularity of zones (e.g., 16 MiB)
Less overprovisioning: ZNS SSDs do not do GC, since the erasure blocks in a zone are always completely invalidated and erased when the zone is reset
ZNS is a more useful abstraction
writes may write to any zone, but each zone must be written sequentially. This thin interface gives applications more control over data placement on flash
since the host has insight into application specifics
ZNS may get better performance
the host is empowered to make application-aware data placement decisions.
ZNS SSD adoption
Research agenda
Improving performance
How can application-level information improve zone management?
software can often make educated guesses about data expiration
How much can filesystem knowledge reduce write amplification?
How should applications interact with zones?
What is the best approach to I/O scheduling with host-driven device management?
How can we best exploit transparent data placement?
Managing limitations
How should hosts manage active zone limts?
Are there workloads that perform worse on ZNS SSDs than on conventional SSDs?
How do ZNS SSDs fit with recent trends to offload I/O tasks from the host to dedicated hardware?
provide some insights when shifting conventional SSDs to ZNS SSDs
Conventional SSD background
flash cells in an erasure block can only be re-written after the block is fully erased
flash cells are hidden behind the traditional block interface
FTL: exposes to the host a flat address space that can be randomly written at a page granularity (4 KiB), similar to HDDs
random writes force the FTL to implement garbage collection to reclaim space from old data that was overwritten in the logical address space
applications have limited control over how the FTL arranges data on the device
ZNS SSDs
a new SSD interface that dominates the conventional block interface in nearly every way
A ZNS SSD is divided into logical regions called zones
the FTL is thinner: coarser-grained address translation (at the erasure block), does not do garbage collection
Hosts control write amplification can make application-aware data placement and I/O scheduling decisions