uFLIP
uFLIP is a flash-device benchmark that varies I/O patterns in space and time while recording individual response times and controlling measurement history.
The CIDR 2009 study defines nine microbenchmarks around request size, alignment, locality, partitioning, address order, parallelism, read/write mix, pauses and bursts. It separates initialization, start-up and running phases, and interference between runs. Its evaluation covered eleven contemporary devices, not a representative sample of current SSDs. [1]
A pause after completion is outside that request's response time but inside the elapsed test interval. The paper reports cases where pauses reduced random-write response times without reducing total workload time. [1] Preserve both clocks when comparing workloads; the linked Hesela example demonstrates the arithmetic without issuing device I/O. [3]
The authors' institute describes uFLIP as a component benchmark for response-time distributions, not a single throughput score. [2]
Relations
Limitations
- Historical device results do not predict modern NVMe performance.
- Writing a whole device was a stated initialization assumption, not proof of a unique physical state.
- Preparation can overwrite data and consume endurance; no hardware reproduction is claimed here.