a64fx utofu put/get レイテンシ測定 experiment

2 ノード間(たぶん 1 tofu 内?)

  PUT Results (One-way latency, half RTT)

  | Size  | Min (µs) | Avg (µs) | Max (µs) | BW (GB/s) |
  |-------|----------|----------|----------|-----------|
  | 8B    | 1.27     | 1.29     | 1.41     | 0.006     |
  | 64B   | 1.27     | 1.29     | 1.38     | 0.050     |
  | 1KB   | 1.53     | 1.56     | 1.70     | 0.655     |
  | 8KB   | 2.72     | 2.76     | 2.94     | 2.97      |
  | 64KB  | 11.75    | 11.83    | 12.60    | 5.54      |
  | 512KB | 84.05    | 84.34    | 85.92    | 6.22      |
  | 1MB   | 166.71   | 167.11   | 168.69   | 6.27      |
  | 8MB   | 1329.05  | 1419.73  | 1543.09  | 5.91      |

  GET Results

  | Size | Min (µs) | Avg (µs) | Max (µs) | BW (GB/s) |
  |------|----------|----------|----------|-----------|
  | 8B   | 2.26     | 2.43     | 2.65     | 0.003     |
  | 64B  | 2.26     | 2.42     | 2.56     | 0.027     |
  | 1KB  | 2.25     | 2.39     | 2.58     | 0.428     |
  | 8KB  | 2.55     | 2.62     | 2.91     | 3.12      |
  | 64KB | 11.59    | 11.61    | 11.74    | 5.65      |
  | 1MB  | 166.52   | 167.65   | 187.52   | 6.25      |
  | 8MB  | 1323.20  | 1335.96  | 1374.54  | 6.28      |

Key Observations

• Small message latency: ~1.3 µs for PUT, ~2.4 µs for GET
• Peak bandwidth: ~6.3 GB/s (approaching TofuD limit of 6.8 GB/s)
• GET overhead: ~1 µs higher due to signaling PUT
• Note: 16MB results show anomaly - needs investigation