microbenchmark · unknown · also Micro-benchmark

Microbenchmark

A microbenchmark measures a small set of operations to isolate a specific performance cost or behavior.

Use focused tests to explain a mechanism, then use a representative application workload or trace to assess its wider effect. Record cache state, request sizes, concurrency and timing boundaries; a macrobenchmark is not automatically realistic either. [1]

For example, sg_read at sg3_utils v1.48 can issue zero-block SCSI READ(10), READ(12) or READ(16) commands to time command overhead. READ(6) does not share that zero-block meaning. A command rate from this test does not establish data-transfer bandwidth or application IOPS. [2]

In Slowfs, researchers delayed ext2 operations through CPU-side execution, not added device I/O service time. CPU-bound compile workloads could conceal that overhead. The injected mechanism matters: a slowdown factor alone cannot predict the response to a physically slower disk. [1]

Relations

related

IOPS, Storage latency, Fault injection

Limitations

• An isolated cost can change when other operations contend for shared resources; combine narrow measurements with tests of the intended workload. [1]

Citations

1. Traeger et al., A Nine Year Study of File System and Storage Benchmarking, FSL-07-01, sections 3.1-3.2 and 12.1-12.2 (accessed 2026-11-10)
2. sg3_utils v1.48: sg_read(8), description and zero-block SCSI READ timing (accessed 2026-11-10)

Further reading

3. Hesela: SCSI command handling, transfer counts and benchmark scope (accessed 2026-11-10)

JSON: /data/terms/microbenchmark.json