SC Windows Storage Benchmarking

This repo now contains two benchmark harnesses:

• storage_perf_windows.ps1: Windows guest benchmark runner using diskspd + Windows perf counters.
• storage_perf.py: Linux guest benchmark runner using fio + iostat + /proc/stat.

If your target is Windows guest behavior, use the PowerShell harness first.

What It Measures

Per workload + iteration:

• Throughput: read_bw_mib_s, write_bw_mib_s, total_bw_mib_s
• IOPS: read_iops, write_iops, total_iops
• Latency: read_lat_mean_us, write_lat_mean_us, plus io_await_* from disk counters
• Queue depth pressure: io_queue_mean, io_queue_p95
• Utilization proxy: io_util_mean_pct, io_util_p95_pct, io_util_max_pct
• CPU iowait:
  • Linux: populated from /proc/stat
  • Windows: not available natively, left blank (cpu_iowait_*), with cpu_busy_* included

Outputs per run:

• metadata.json
• workloads.json
• summary.csv
• results.csv (same raw rows as summary.csv, for easier downstream naming)
• dashboard.html (single-run web dashboard)
• dashboard.pdf (single-run PDF report, best-effort via headless Edge/Chrome)
• raw/ logs and counter snapshots

Windows Guest (DiskSpd)

Prereqs

• Windows guest with admin shell
• diskspd.exe on PATH, or let the harness auto-install it (default behavior)
• PowerShell 5.1+ or 7+

If diskspd is missing, the harness now tries:

1. Existing known folders (.\tools\diskspd, local app data install dir, common Program Files paths)
2. winget package install
3. Official ZIP download/install (https://aka.ms/diskspd) into local install dir

Run baseline (example: virtio-scsi)

Set-Location C:\bench\storage-perf
.\storage_perf_windows.ps1 `
  -Label virtio-scsi `
  -TargetPath C:\bench\testfile.dat `
  -Size 20G `
  -DurationSeconds 60 `
  -WarmupSeconds 10 `
  -Repeat 3 `
  -OutputDir .\results

Then switch VM disk bus to virtio-blk and run:

.\storage_perf_windows.ps1 `
  -Label virtio-blk `
  -TargetPath C:\bench\testfile.dat `
  -Size 20G `
  -DurationSeconds 60 `
  -WarmupSeconds 10 `
  -Repeat 3 `
  -OutputDir .\results

Compare two runs

.\storage_perf_windows.ps1 -Compare `
  -A .\results\virtio-scsi-20260211-120000 `
  -B .\results\virtio-blk-20260211-123000 `
  -CompareOutput .\compare-scsi-vs-blk.csv

Compare mode also writes a matching HTML dashboard next to the CSV (same name, .html extension). If browser-based PDF export is available, compare mode also writes a matching .pdf.

Optional knobs

• -DiskInstance "0 C:" to pin which PhysicalDisk(*) instance gets summarized
• -WorkloadsJson .\my_workloads.json to override workload matrix
• -ExtraDiskSpdArg "-h" (or repeated) to pass extra diskspd flags
• -AutoInstallDiskSpd $false to disable auto-install attempts
• -DiskSpdInstallDir "$env:LOCALAPPDATA\diskspd" to control local auto-install location
• -GeneratePdf $false to skip PDF generation
• -PdfBrowserPath "C:\Program Files\Microsoft\Edge\Application\msedge.exe" to force a specific browser binary for PDF export

Workloads JSON format (Windows)

[
  { "name": "seq_read_1m_qd32", "rw": "read", "bs": "1M", "o": 32, "t": 1 },
  { "name": "rand_rw_4k_70r_qd32", "rw": "randrw", "bs": "4K", "o": 32, "t": 1, "rwmixread": 70 }
]

Supported rw: read, write, randread, randwrite, randrw. Use t (threads) and o (outstanding I/O per thread) to explore low-vs-high concurrency regimes where backend differences usually appear.

Linux Guest (fio) Optional

If you also want Linux-side confirmation:

python3 storage_perf.py run --label virtio-scsi --filename /mnt/bench/testfile --size 20G
python3 storage_perf.py run --label virtio-blk --filename /mnt/bench/testfile --size 20G
python3 storage_perf.py compare --a results/virtio-scsi-... --b results/virtio-blk-...

Experiment Hygiene (Important for Fair A/B)

• Keep VM CPU/memory topology identical between virtio-scsi and virtio-blk.
• Keep backing storage identical (same host disk, cache mode, qcow/raw format).
• Reboot between variants, or at least clear caches and let background IO settle.
• Use same workload file size and test path.
• Keep test duration long enough for tail latency to stabilize (60s+ is a good start).
• Run at least 3 iterations and compare means + variability.