How to Use fio to Measure Disk IOPS, Throughput, and Latency on Ubuntu

Introduction

fio is a flexible I/O workload generator for measuring storage performance (IOPS, bandwidth, latency). It can issue sequential or random reads/writes, run multiple threads, and simulate heavy load.

Test environment

Ubuntu 20.04.5 LTS, kernel 5.4.0-125-generic. Install fio:

# sudo apt update
# sudo apt install fio
# fio --version
fio-3.16

IOPS / bandwidth benchmark

Run a mixed random read/write test on block device /dev/vdd with four jobs, 1 MiB block size, queue depth 32, and 100 GiB per job for 120 seconds.

fio \
  --name=iops-test \
  --filename=/dev/vdd \
  --rw=randrw \
  --bs=1M \
  --numjobs=4 \
  --iodepth=32 \
  --size=100G \
  --runtime=120 \
  --direct=1 \
  --ioengine=libaio \
  --group_reporting

Key parameters: --filename : target block device --rw=randrw : mixed random read/write --bs=1M : I/O size --numjobs=4 : concurrent jobs (threads) --iodepth=32 : queue depth --size=100G : data per job --runtime=120 : test duration (seconds) --direct=1 : bypass page cache --ioengine=libaio : Linux asynchronous I/O --group_reporting : aggregate output

Important output columns:

BW : bandwidth (MiB/s or MB/s)

IOPS : operations per second

slat : submission latency (time from request submission to kernel entry)

clat : completion latency (time kernel spends processing the request)

Latency‑focused benchmark

To measure raw latency, use a small block size, a single job, and queue depth 1, which disables queue‑level optimizations.

fio \
  --name=latency-test \
  --filename=/dev/vdd \
  --rw=randread \
  --bs=4k \
  --numjobs=1 \
  --iodepth=1 \
  --size=10G \
  --runtime=60 \
  --direct=1 \
  --ioengine=libaio \
  --group_reporting

Focus on slat and clat values to assess the base response time of the device.

Interpreting results

For the IOPS test, examine the BW and IOPS columns to evaluate throughput and operation rate. For the latency test, review slat and clat percentiles to understand typical and worst‑case response times. Adjust --bs, --iodepth, and --numjobs to target different performance aspects (bandwidth vs. latency).