Glossary

IOPS

IOPS, input/output operations per second, counts how many separate read and write requests a storage system completes each second. It measures requests rather than data volume, so a figure means little without the size of each operation and the latency at which the rate was reached.

How an IOPS figure is expressed

Every request counts as one operation, whether it asks for 4 KB or 4 MB. A complete figure therefore states the operation size, the read and write mix, whether access was random or sequential, how many requests were outstanding at once, and the latency measured at that rate. Remove any of those and two figures cannot be compared.

Operation size changes what the number means. A system completing 100,000 operations of 4 KB each moves about 400 MB per second. One completing 1,000 operations of 4 MB moves about 4 GB per second with a hundredth of the IOPS. The first is the measure that matters for a database; the second is a throughput result.

What moves the number

  • Media. A hard drive manages roughly 100 to 200 scattered operations per second, because a mechanical arm has to move for each one. An SSD or NVMe drive handles tens of thousands or more.
  • Access pattern. Scattered requests are far harder on disks than ordered ones, a gap covered under sequential vs random I/O.
  • Write share. Writes carry protection overhead, so a write-heavy mix lowers the rate the same hardware reaches.
  • Concurrency. More requests in flight let a system overlap work and raise IOPS, at the cost of each request waiting longer, so latency belongs beside every figure.
  • Cache hits. Requests answered from a flash cache never reach the slower media.

Why does capacity not buy IOPS?

A hard drive delivers about the same number of random operations whether it holds 4 TB or 20 TB. Moving to larger drives in the same chassis adds space while the operations available for each terabyte fall. Sizing in IOPS per terabyte exposes this: archive and backup data need very little, while active databases and virtual machine datastores need a lot. Flash and hard disk are often combined in one system for exactly this split.

Reading a datasheet value

A published maximum is normally a best case: small operations, a favorable mix, a short run and a deep queue of waiting requests. A workload with larger operations, more writes or fewer requests in flight lands well below it. The useful comparison is with the workload's own peak, measured at its busiest hour, at the latency the application can tolerate. A figure from a vendor test that matches neither the operation size nor the mix of that peak is a ceiling, not a forecast.

Frequently asked questions

Is a higher IOPS number always better?

No. A high figure reached with long waits per request, or at an operation size the application never uses, can still give a poor experience. The number is useful only next to operation size and latency.

Do IOPS matter for object storage?

For large objects, throughput matters more. For many small objects and metadata lookups, the operation rate becomes the limit and IOPS per terabyte becomes a real sizing input.

How many IOPS does a workload need?

The reliable answer comes from measuring the existing workload at its busiest hour, with its own operation sizes, and sizing to that peak plus headroom for growth.