GitHub - haydenjames/VPS-Disk-Latency-Bench: Latency-focused disk benchmarks for Linux VPS environments.

4 min read Original article ↗

Latency-focused disk benchmarks for VPS environments.

This script runs small-block mixed random I/O at low queue depths and light concurrency, then prints a formatted results table with color-coded latency assessments.

VPS Disk Latency Bench Screenshot

Why latency, not IOPS?

Most VPS workloads (web servers, databases, applications) operate at queue depth 1-2 with single-threaded I/O. High IOPS numbers from synthetic benchmarks at qd32+ don't reflect real-world performance.

Why VPS focused?

This script works on any Linux system: dedicated servers, cloud instances, homelab boxes, local workstations. The VPS framing is about audience, not technical scope. VPS buyers are the ones most often misled by inflated IOPS numbers from providers running synthetic qd32 benchmarks on marketing pages. But anyone evaluating storage latency under realistic workloads will find the results useful. It's just fio with sensible parameters.

This script tests what actually matters: p99.9 latency at low queue depth.

Read more: VPS IOPS vs. Latency: Why NVMe Benchmarks Lie

Features

  • Formatted table output with aligned columns
  • Color-coded p99.9 latency (green/yellow/red thresholds)
  • Key metrics summary highlighting the most important qd1 results
  • Pass/fail assessment against 0.3ms NVMe threshold
  • Progress indicator during test execution
  • JSON + TXT output for both human reading and scripting

Requirements

Required:

  • fio

Standard tools (present on most Linux systems):

  • dd, awk, sed, date, hostname

Optional:

  • bc (used for floating-point comparisons; falls back to awk)
  • tput (for colorized output, if supported)
  • jq (only needed if you want to manually inspect the raw JSON output)

Quick Start

# Download
curl -O https://raw.githubusercontent.com/haydenjames/VPS-Disk-Latency-Bench/main/bench-io.sh
chmod +x bench-io.sh
./bench-io.sh # Run with defaults (2GB file, 30s per test)
# Quick test (smaller file, shorter runtime)
FILE_SIZE_GB=1 RUNTIME_SEC=10 ./bench-io.sh

Sample Output

╔═══════════════════════════════════════════════════════════════════════════╗
║             VPS Disk Latency Benchmark - Latency-Focused Testing          ║
╚═══════════════════════════════════════════════════════════════════════════╝

Test Name                        │    Read IOPS   Avg (ms)      p99.9 │   Write IOPS   Avg (ms)      p99.9
─────────────────────────────────┼─────────────────────────────────────┼─────────────────────────────────────
randrw_70read_4k_qd1_jobs1       │        9,013      0.089      0.157 │        3,882      0.036      0.089
randrw_70read_4k_qd1_jobs2       │       17,238      0.091      0.214 │        7,397      0.041      0.098
...

═══════════════════════════════════════════════════════════════════════════════
                              KEY METRICS (qd1_jobs1)
═══════════════════════════════════════════════════════════════════════════════

  4K Random Read/Write (70/30) at Queue Depth 1:

    Read  p99.9 Latency:  0.157 ms  ✓ Excellent (NVMe-class)
    Write p99.9 Latency:  0.089 ms  ✓ Excellent (NVMe-class)

  Latency Thresholds:
    < 0.3ms  = Excellent (true NVMe-class performance)
    < 0.5ms  = Acceptable
    >= 0.5ms = Poor (likely throttled, shared, or not true NVMe)

Configuration

Override defaults via environment variables:

Variable Default Description
FILE_DIR /var/tmp Directory for test file
FILE_SIZE_GB 2 Test file size in GiB
RUNTIME_SEC 30 Runtime per test in seconds
WARMUP_SEC 5 Warmup time before each test
IODEPTHS 1 2 4 Queue depths to test
BS_LIST 4k 8k Block sizes to test
MIX_LIST 70 50 Read percentages (70 = 70% read, 30% write)
JOBS_LIST 1 2 Number of parallel jobs
ENGINE libaio fio I/O engine
OUTPUT_DIR ./bench_out Output directory

Examples

# Minimal test (fastest)
FILE_SIZE_GB=1 RUNTIME_SEC=5 IODEPTHS="1" BS_LIST="4k" MIX_LIST="70" JOBS_LIST="1" ./bench-io.sh

# Full test with larger file
FILE_SIZE_GB=4 RUNTIME_SEC=60 ./bench-io.sh

# Custom output directory
OUTPUT_DIR=/tmp/benchmarks ./bench-io.sh

Output Files

The script creates two files in ./bench_out/ (or $OUTPUT_DIR):

File Description
fio-<hostname>-<timestamp>.txt Formatted summary with tables, key metrics, and raw data
fio-<hostname>-<timestamp>.json Raw fio JSON output for programmatic analysis

What to Look For

Focus on p99.9 latency in the 4k_qd1_jobs1 tests:

p99.9 Latency Assessment
< 0.3ms ✅ Excellent - True NVMe-class performance
< 0.5ms ⚠️ Acceptable - Reasonable for most workloads
≥ 0.5ms ❌ Poor - Likely throttled, noisy neighbors, or not true NVMe

Warning signs:

  • IOPS that flatline at round numbers (e.g., 20,000) suggest rate limiting
  • High IOPS with high latency indicates oversubscribed storage
  • Erratic latency across tests suggests noisy neighbors

Changelog

v2.0.0 (2026-02-03)

  • New: Formatted table output with Unicode box drawing
  • New: Color-coded p99.9 latency (green/yellow/red)
  • New: Key metrics summary section
  • New: Pass/fail assessment against thresholds
  • New: Progress indicator during tests
  • New: Improved TXT output with multiple sections
  • Fixed: Compatibility with various fio versions
  • Fixed: Robust JSON parsing

v1.0.0

  • Initial release

License

MIT

Author

Hayden James - Linux systems analyst.