fio_bench_v1.sh

#!/bin/bash
# =============================================================================
# 🏎️  NFS STORAGE BENCHMARK — Audio File I/O Profile (PyTorch-Realistic)
# =============================================================================
#
# PURPOSE:
#   Benchmarks an NFS (or similar networked) filesystem for ML training pipelines
#   that consume large numbers of small WAV/FLAC audio files using PyTorch-style
#   dataloaders.
#
# MODELLED WORKLOAD:
#   - Many small files (~3 MB each)
#   - Random file access (shuffle every epoch)
#   - Synchronous reads (iodepth=1)
#   - Parallelism from worker count, not async I/O
#
#   This closely matches PyTorch Dataset.__getitem__ behavior and exposes
#   metadata latency (open/stat/close), which is the dominant NFS bottleneck.
#
# MULTI-NODE USAGE:
#   Run simultaneously on all nodes to measure aggregate pressure:
#     pdsh -w node[01-08] '/path/to/storage_bench.sh /mnt/shared/bench_$(hostname)'
#
# DEPENDENCIES: fio, python3
# =============================================================================

set -euo pipefail

# ── Configuration ────────────────────────────────────────────────────────────
DIR="${1:-/data_vast/bench_$(hostname)}"
JOBS=32            # Parallel workers (match PyTorch num_workers)
RUNTIME=180s       # Duration per test
RAMP=5s            # Warm-up period
NRFILES=64         # Files per worker
FILESIZE=3m        # Avg WAV/FLAC size
TOTAL_SIZE=192m    # NRFILES × FILESIZE

# ── Preflight Checks ─────────────────────────────────────────────────────────
command -v fio >/dev/null 2>&1 || {
  echo "ERROR: fio not found. Install with: apt install fio (or yum install fio)"
  exit 1
}
command -v python3 >/dev/null 2>&1 || {
  echo "ERROR: python3 not found."
  exit 1
}

mkdir -p "$DIR"
trap 'echo ""; echo "Cleaning up test files..."; rm -f "$DIR"/bench.* 2>/dev/null; echo "Done."' EXIT

# ── Header ───────────────────────────────────────────────────────────────────
TOTAL_FILES=$((JOBS * NRFILES))
TOTAL_GB=$(python3 -c "print(f'{$TOTAL_FILES * 3 / 1024:.1f}')")

echo "========================================================================"
echo " 🏎️  NFS STORAGE BENCHMARK — Audio File I/O Profile"
echo " 🖥️  Node:       $(hostname)"
echo " 📁  Target:     $DIR"
echo " ⚙️  Config:     $JOBS workers × $NRFILES files × $FILESIZE each"
echo " 📊  Footprint:  $TOTAL_FILES files, ~${TOTAL_GB} GB"
echo " ⏱️  Runtime:    $RUNTIME (+ ${RAMP} ramp)"
echo " 📌  Model:      PyTorch-style synchronous I/O (iodepth=1)"
echo "========================================================================"
echo ""

# ── FIO Runner ───────────────────────────────────────────────────────────────
run_fio() {
  local label="$1" rw="$2" bs="$3"
  shift 3
  local extra_args=("$@")

  printf "  %-30s ... " "$label"

  result=$(fio --name=bench --group_reporting --time_based=1 \
    --runtime="$RUNTIME" --ramp_time="$RAMP" \
    --size="$TOTAL_SIZE" --filesize="$FILESIZE" \
    --nrfiles="$NRFILES" --file_service_type=random \
    --numjobs="$JOBS" \
    --ioengine=libaio \
    --iodepth=1 \
    --rw="$rw" --bs="$bs" \
    --directory="$DIR" \
    --thread \
    --fallocate=none \
    --allow_file_create=1 \
    "${extra_args[@]}" \
    --output-format=json 2>/dev/null) || {
      echo "❌ FAILED (check mount/permissions at $DIR)"
      return
    }

  stats=$(echo "$result" | python3 -c "
import sys, json
try:
    j = json.load(sys.stdin)['jobs'][0]
    res = j['read'] if j['read']['bw_bytes'] > 0 else j['write']
    bw   = res['bw_bytes']
    iops = res['iops']
    p99  = res['clat_ns']['percentile'].get('99.000000', 0)
    print(f'{bw} {iops} {p99}')
except Exception:
    print('0 0 0')
")

  read -r bw iops p99 <<< "$stats"

  if [ "$bw" = "0" ] && [ "$iops" = "0" ]; then
    echo "⚠️  No data (test too short or stalled mount)"
    return
  fi

  bw_mb=$(python3 -c "print(f'{$bw / 1e6:.0f}')")
  iops_k=$(python3 -c "print(f'{$iops / 1e3:.1f}')")
  p99_ms=$(python3 -c "print(f'{$p99 / 1e6:.1f}')")

  printf "✅ %6s MB/s | %7s K IOPS | p99: %s ms\n" "$bw_mb" "$iops_k" "$p99_ms"
}

# =============================================================================
# Small-File Write (Direct I/O)
# =============================================================================
# Models preprocessing pipelines writing many small audio files to shared storage.
echo "── Small-File Write (${TOTAL_FILES} × ${FILESIZE}, direct I/O) ───────────────"
run_fio "Rand Write 1M (iodepth=1)" randwrite 1m --direct=1
echo ""

# =============================================================================
# Small-File Read (Direct I/O)
# =============================================================================
# Models shuffled PyTorch dataloader reads with cold page cache.
echo "── Small-File Read (${TOTAL_FILES} × ${FILESIZE}, direct I/O) ────────────────"
run_fio "Rand Read  1M (iodepth=1)" randread 1m --direct=1
echo ""

# =============================================================================
echo "========================================================================"
echo " ✅  Benchmark Complete"
echo ""
echo " INTERPRETATION:"
echo ""
echo "   Read p99 latency is the primary signal."
echo ""
echo "     p99 < 5 ms   → Excellent (safe for multi-GPU training)"
echo "     5–10 ms      → Acceptable but watch scaling"
echo "     10–20 ms     → Intermittent GPU stalls likely"
echo "     >20 ms       → Training throughput collapse at scale"
echo ""
echo "   If results are poor, consider:"
echo "     - Increasing NFS rsize/wsize (e.g. 1MB)"
echo "     - Multiple client IPs / nconnect"
echo "     - Packing files into tar/WebDataset shards"
echo "========================================================================"