Glossary

Flash cache

A flash cache is a layer of flash that holds copies of frequently read data in front of slower storage, usually hard disks. Requests found in the cache are answered at flash speed and the rest go to disk, so the benefit depends on how often the cache can answer rather than on the flash itself.

Hits, misses and the average

A request found in the cache is a hit; one that is not is a miss and goes to disk. A disk read takes roughly a hundred times longer than a flash read, so a small share of misses dominates average latency. Falling from 99 percent hits to 95 percent roughly triples the average read time, even though the cache still answers nearly every request. Two systems with identical flash can therefore behave very differently, because their workloads revisit data at different rates. The hit rate reported by the system over a normal working week is the figure worth watching; the size of the flash on the datasheet is not. A falling hit rate over months usually means the active data has outgrown the cache.

Which workloads gain little?

  • Read-once work: migrations, full backup reads, one-time analytics sweeps and training passes over a large data set fill the cache with data nobody asks for again.
  • Working sets larger than the cache, where new data evicts old before it is reused.
  • Write-heavy work, since writes still have to reach durable media.
  • Anything immediately after a restart: a new or restarted cache starts empty and needs hours of normal traffic before its hit rate settles.

Cache or tier

A cache keeps copies; the original stays on disk, so losing a read cache costs speed but no data. A write cache is different, because it can hold data not yet flushed, and it has to be protected against power loss. Storage tiering moves the data itself between layers on a schedule or policy, so the fast layer holds the only copy until data moves back. Hybrid flash arrays use one approach or both to put a small amount of flash in front of a lot of disk. A cache reacts within minutes to a change in what is being read, while a tier reacts only when its policy next runs, so caching suits workloads whose active data shifts often and tiering suits data whose activity fades predictably with age.

Sizing to the working set

The size that matters is the part of the data read repeatedly, not the whole data set. A cache holding most of that active portion delivers most of the gain; a cache larger than it adds little. In object storage, flash is often used for metadata and small objects, which generate many small requests measured in IOPS, while large objects stay on disk. The data that stays in heavy use is what the glossary calls hot storage.