Glossary
Sequential vs random I/O
Sequential vs random I/O describes the two basic storage access patterns. Sequential I/O reads or writes data in order, each request continuing where the last ended. Random I/O jumps between unrelated locations. The same hardware can perform very differently under each, so the pattern decides which performance measure is meaningful.
The two patterns compared
| Sequential | Random | |
|---|---|---|
| Typical workloads | Backup streams, video, large file copies, analytics scans | Databases, virtual machines, small files, metadata and index updates |
| Usual request size | Large | Small |
| On hard drives | Close to the drive's full streaming rate | A small fraction of it, limited by arm movement |
| On flash | Fast | Also fast; the gap is much narrower |
| Measure that matters | Throughput | IOPS and latency |
Why disks feel the difference most
A hard drive reads ordered data as the platter turns beneath the head. A scattered request first needs the arm to move and the right sector to arrive, so a drive that streams hundreds of megabytes per second completes only 100 to 200 random operations per second. Flash has no moving parts, which is one reason random workloads moved to flash first while large sequential data stayed on disk, a split covered under flash storage vs HDD.
Mixed patterns and concurrency
Few systems see a pure pattern. Twenty backup jobs that are each perfectly sequential arrive at shared drives interleaved, and the drives see something close to random access unless the storage system groups the writes. Concurrency matters as well: a benchmark with many requests in flight lets a system reorder and overlap work, while a database issuing a few at a time cannot benefit in the same way. A figure measured under one pattern and queue depth says little about another, so a datasheet that does not state both cannot be matched to a workload. Measuring the real mix, including how many requests each application keeps in flight, is the only reliable way to know which pattern dominates.
What does object storage favor?
Object storage writes objects whole and usually reads them whole, so large objects produce sequential work that suits high-capacity disk. Small objects and the metadata that tracks every object produce random work. Designs that hold metadata on flash, or put a flash cache in front of disk, serve both patterns from one system.
Frequently asked questions
Is sequential I/O always faster?
On hard drives, by a wide margin. On flash the gap is much smaller, and the difference shows mainly in how many operations per second a device sustains.
What is queue depth?
The number of requests outstanding at the storage at the same time. Deeper queues let a system overlap work and raise IOPS, but each request may wait longer.