Hybrid flash array
A hybrid flash array is a storage system that combines flash with hard-disk capacity and uses caching or tiering to keep frequently accessed data on flash.
Cache safety depends on write policy. A clean cached copy can be discarded when a valid backing copy remains. A dirty write-back block can contain newer data than the origin; losing it can lose updates. Cache data and metadata need protection appropriate to the failure model. [4] [5]
Linux dm-cache uses origin, cache and metadata devices. Its kernel target defaults to writeback; LVM documents writethrough as its default, configurable by the administrator. Check the active mode rather than inferring it from the word cache. Writethrough completes writes to cached blocks on both devices, but downstream volatile caches still matter. [2] [5]
Before removing dirty cached data, write it back to the origin and verify completion through the supported management procedure. dm-cache requires a clean cache for passthrough and rejects shrinking over dirty blocks. A persistent cache can still be dirty relative to its backing store. After a crash, dm-cache assumes all cache blocks dirty because dirty-state changes are not persisted individually; this is not a count of lost writes. [4]
One scheduled-tiering design moves 1 GB units under a per-volume policy, ranks activity hourly with a roughly 24-hour half-life and reserves 10% of higher-tier space for new allocations. Those placement rules describe that design, not every hybrid array. [3]
In Linux 6.12 dm-cache, consistent recovery metadata does not ensure every recent write survives. Metadata commits accompany FLUSH/FUA and otherwise occur periodically. Writes not covered by a completed durability operation may be lost on power failure. A completed flush is not evidence that all dirty blocks reached the origin. [4] [6]
Relations
- related
- Solid-state drive (SSD), Hard disk drive (HDD), Hot, warm, and cold storage tiers, All-flash array, Write-back cache, Logical unit number (LUN), SSHD (hybrid drive), Write-through cache, Storage cache flush, Force Unit Access, Storage array, Automated storage tiering
Advantages
- The cited scheduled-tiering design can combine caching for short-term access changes with relocation for longer-term placement. [3]
- In dm-cache, the origin covers every logical block, even when its version is stale in writeback mode. [2]
Limitations
- Transfer granularity is implementation-specific: dm-cache accepts 32 KiB to 1 GiB chunks, while the cited tiering design specifies a 1 GB unit. [2, 3]
- A 1 TiB origin divided into 512 KiB blocks has 2,097,152 addressable blocks; this is not a metadata-size estimate or a count of resident cache entries. [2]
Citations
- Linux kernel documentation: dm-cache (device-mapper cache target)
- QSAN. Auto Tiering white paper, February 2026.
- Linux 6.12 documentation: dm-cache operating modes, cleaner and metadata
- LVM2 lvmcache(7): cache modes, chunk size, migration threshold and RAID1 cache devices
- Linux 6.12: Explicit volatile write back cache control, flush/FUA completion and remapping drivers
Further reading
Sources & caveats
- QSAN Auto Tiering, February 2026, sections 2.1-2.3 supplies the placement parameters and cache/tiering comparison; its disposable-copy description does not apply to every write-back cache. [3]
- Linux 6.12 supplies the pinned cache-state and removal rules; LVM2 documents management-layer defaults and optional RAID1 for cache data and metadata. [4] [5]
- The kernel's documented 2048-sector migration threshold is 2048 x 512 = 1,048,576 bytes = 1 MiB; LVM can raise it to at least eight chunks. [4] [5]
- Arithmetic: 1 TiB / 512 KiB = 2,097,152 origin blocks. Actual metadata depends on cache size, format and implementation.
- Source comparison, not a hardware-failure experiment or a measured array speed-up.
- Linux 6.12 documents metadata commits every second without FLUSH/FUA. This is a documented cadence, not an application persistence deadline. [4]
- The origin-cleaning distinction is an inference from the documented writeback and cleaner contracts; flush/FUA propagation and completion remain separate requirements. [4] [6]