Benchmarking Cloud VPS Disks with FIO: Storage Performance

Accurately benchmarking disk performance on cloud servers is often misunderstood. Simple tests using the dd command (such as dd if=/dev/zero of=test.bin) fail to simulate real workloads: dd writes large sequential streams of zeroed memory that are trivially compressed by modern storage controllers and absorbed entirely into the RAM page cache. The resulting figures report multi-gigabyte numbers that bear zero resemblance to production database performance.

Flexible I/O Tester (FIO) is the enterprise industry standard for storage benchmarking. FIO bypasses the page cache with direct I/O (O_DIRECT), generates cryptographically random payloads that defeat storage compression tricks, simulates multi-threaded database access patterns, and reports granular 99th and 99.9th percentile latency distributions.

Expand

Authoring an Enterprise FIO Benchmark Job

Create a comprehensive multi-test job configuration file at benchmark.fio:

# benchmark.fio
[global]
ioengine=io_uring
direct=1
runtime=45
time_based=1
filename=/tmp/fio_benchmark.tmp
size=8G
group_reporting=1

# Test 1: High concurrency 4K random read (OLTP reads)
[4k-random-read]
rw=randread
bs=4k
iodepth=64
numjobs=4

# Test 2: Synchronous 4K random write (Database WAL flush)
[4k-random-write]
rw=randwrite
bs=4k
iodepth=16
numjobs=2
stonewall

# Test 3: Large block sequential throughput
[1m-sequential-throughput]
rw=read
bs=1M
iodepth=16
numjobs=1
stonewall

# Test 4: Realistic 70% read / 30% write mixed profile
[70-30-mixed-database]
rw=randrw
rwmixread=70
bs=8k
iodepth=32
numjobs=4
stonewall

Executing the Test and Interpreting Output

Execute the job and inspect both raw text summaries and JSON output:

# Run FIO with output redirection
sudo fio benchmark.fio --output=benchmark-results.txt

# Inspect IOPS and 99th percentile latency for the mixed database test
grep -A 10 "70-30-mixed-database" benchmark-results.txt | grep -E "IOPS|clat percentiles"

Critical metrics to evaluate:

• clat (Completion Latency): The duration elapsed between submitting the I/O to the kernel and receiving hardware completion.
• Percentiles (99.00th, 99.90th): If p99.9 exceeds 15ms during mixed workloads, your storage backend is experiencing queue saturation.

Storage Performance Video

Explore RAID arrays, disk striping, and storage throughput dynamics in this overview:

Avoiding Drive Wear during Benchmarks

When testing storage on cloud instances, always remove benchmark scratch files immediately following job completion with rm -f /tmp/fio_benchmark.tmp. For flash-based SSDs, executing fstrim -v / after extensive benchmarking informs the controller to reclaim discarded flash blocks via background garbage collection.

Benchmark consistent High-Speed SSD performance on your own instance at netbayhosts.in.