Don't Be a Blockhead: Zoned Namespaces Make Work on Conventional SSDs Obsolete

1. Summary

Motivation of this paper

• Motivation

  • Industry has standardized and is adopting Zoned Namespaces (ZNS) SSDs, which offer a new storage interface that dominates conventional SSDs

    • argue for an immediate and complete shift in research to ZNS SSDs and discuss research directions

ZNS SSD

• 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

      • system resources can be allocated according to an application's needs, rather than being fixed for all applications

  • 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

      • more control over write amplification

    • since the host has insight into application specifics

      • can manage GC and other data management tasks better than an on-board FTL

  • ZNS may get better performance

    • the host is empowered to make application-aware data placement decisions.

      • can control what data will be erased together by writing it to the same zone

  • ZNS SSD adoption

    • Btrfs and ext4 are close to running natively on ZNS SSDs
    • Linux provides the dm-zoned device mapper which emulates a standard block device on top of any ZNS SSD
    • The SPDK of developing kernel-bypass storage applications, supports ZNS as of version 20.10

• Research agenda

  • Improving performance

    • How can application-level information improve zone management?

      • software can often make educated guesses about data expiration

        • pages in the same file are more likely to get deleted or overwritten together than pages in different files

    • How much can filesystem knowledge reduce write amplification?

    • How should applications interact with zones?

      • raw zoned storage access offers the most control over I/O and data placement
      • Interacting with zoned flash devices in a way that maximizes I/O performance, usability, and device endurance has not been fully explored

    • What is the best approach to I/O scheduling with host-driven device management?

      • with ZNS SSDs, the performance of individual zones is independent of one another, beyond sharing bandwidth.

    • How can we best exploit transparent data placement?

  • Managing limitations

    • How should hosts manage active zone limts?

      • current ZNS SSDs limit the number of zones that can be writable at the same time
      • it becomes a problem when a ZNS SSD is divided among applications that bypass the kernel

    • Are there workloads that perform worse on ZNS SSDs than on conventional SSDs?

      • zone's write pointer can suffer from lock contention. It is a problem for multi-writer workloads (writes are concentrated in a single zone)

    • How do ZNS SSDs fit with recent trends to offload I/O tasks from the host to dedicated hardware?

2. Strength (Contributions of the paper)

• provide some insights when shifting conventional SSDs to ZNS SSDs

  • includes some potential research topics

3. Weakness (Limitations of the paper)

4. Some Insights (Future work)

• 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

      • write amplification

  • 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

    • matching how data is written to physical flash (abstracting the messy details of the flash hardware)

  • A ZNS SSD is divided into logical regions called zones

    • writes can go to any zone but must be sequential within a zone

  • the FTL is thinner: coarser-grained address translation (at the erasure block), does not do garbage collection

    • performance variability and other impacts of GC are eliminated

  • Hosts control write amplification can make application-aware data placement and I/O scheduling decisions