DACP

Check-dam analysis: CONET versus FTP download-first processing

2026-11-08T12:21:10Z

Three data nodes, Union and CONET run on 39.106.98.67; FTP downloads and centralized computation run on the local client. Data are synthetic, without bandwidth shaping. Client-to-server traffic uses the public WAN; the three centers remain colocated, not geographically distributed.

Environment

{
  "client": {
    "platform": "Linux-6.8.0-85-generic-x86_64-with-glibc2.39",
    "logical_cpus": 16,
    "network": "Public WAN: local client to 39.106.98.67; three remote centers colocated; no shaping; direct HTTP and FTP"
  },
  "remote": {
    "host": "39.106.98.67",
    "logical_cpus": 8,
    "memory_gib": 15,
    "os": "Ubuntu 26.04 x86_64",
    "java": "21.0.12.1",
    "runtime_image": "xstream/check-dams-demo:1.0",
    "runtime_image_id": "sha256:ee77fe0eed057308945ef0406e5fcccacb829d1de3cb00dccd5566a6db90db3e",
    "jvm_heap": "64m–512m per service",
    "other_services": "Existing unrelated applications remain running",
    "layout": "Three source nodes, Union, CONET and FTP share one remote host"
  }
}

Client and server hardware differ, and unrelated services share the remote server. This measures entire deployment paths, not a protocol or engine in isolation. CONET HTTP control and final results cross the WAN; inter-node transfers remain local to the remote host.

One task, two execution paths

flowchart LR
    A[Terrain center] --> AT[Terrain metrics and susceptible channels]
    B[Imagery center] --> BT[YOLO candidate detection]
    C[Hydrology center] --> U[Hydrology node: spatial overlap and map]
    AT --> U
    BT --> U
    Planner[Union planning] -.-> AT
    Planner -.-> BT
    Planner -.-> U
    U --> Result[Validated analysis result]
    A --> FTP[FTP: download all three original bundles]
    B --> FTP
    C --> FTP
    FTP --> Local[Local XStream: identical Job]
    Local --> Result

The diagram shows the current execution paths: Union plans, and data nodes execute their assigned steps. Exact placement is listed below.

Inputs and result equality

Source data total 1,018,060,800 bytes (1.018 GB, 970.9 MiB). The analysis uses the default first-tile polygon and confidence 0.35. All original bundles are retained, without cropping for the local path.

{
  "region": {
    "type": "Polygon",
    "coordinates": [
      [
        [
          109.1107897841093,
          36.05322366614358
        ],
        [
          109.20694875299046,
          36.05322366614358
        ],
        [
          109.20694875299046,
          36.131268701939064
        ],
        [
          109.1107897841093,
          36.131268701939064
        ],
        [
          109.1107897841093,
          36.05322366614358
        ]
      ]
    ]
  },
  "min_confidence": 0.35
}

All runs validate every field, type, list length and order. Integers, strings and selection decisions match exactly; cross-CPU floating values use explicit absolute tolerances: longitude/latitude 1e-9 degrees, projected/pixel positions 1e-4 native units, confidence 1e-6, and other floats 1e-8. Original JSON hashes and difference paths are retained; results are not claimed to be byte-identical. Every run has 5 candidates, 3 valid sites, 4 rejections and 361 inference windows.

Measured time

Each path has 1 warmups and 3 measured runs, interleaved with a fixed random seed. Values are medians and min–max ranges, not confidence intervals.

Path Median end-to-end Min–max
CONET / Union 16.464 s 13.874–18.168 s
FTP + local XStream 99.217 s 95.505–106.121 s

Centralized / CONET time ratio: 6.03×. Above 1 means CONET is faster; below 1 means centralized processing is faster.

End-to-end time

Centralized stages

Stage Median time
Parallel FTP download, hash, write and close 73.935 s
Local engine startup, computation and output write 25.069 s
Read and validate full analysis result 0.001 s

Independent stage medians need not sum to the median total.

Centralized stage times

Actual placement and remote payload

Center Steps
Terrain center terrain_data, terrain_metrics, susceptible_channels
Imagery center imagery, candidate_sites
Hydrology center channel_grids, spatial_overlap, distribution_map

Median observed remote payload across CONET node output ports is 40,147 bytes, versus 1,018,060,800 FTP source bytes per centralized run. In this batch overlap runs on the channel-grid node, avoiding a full channel-grid transfer. Placement differs from the previous local batch, so its traffic cannot be assumed identical. These are not TCP/TLS wire-byte measurements. The final SVG is 14,816 bytes and is not the traffic total of the collaboration.

Measurement boundaries

Individual trials and reproduction

Measured repeat CONET s FTP centralized s
1 18.168 106.121
2 16.464 95.505
3 13.874 99.217

Of 3 paired measured runs, CONET is faster in 3 and centralized processing in 0. The small sample does not establish a stable speedup or statistical significance.

Sanitized raw measurements

Reproduction

The CONET benchmark tools provide isolated remote deployment, FTP downloads and local execution of the same Job. Prepare data, images and operators first; keep connection configuration private. Run these commands in the CONET checkout:

python benchmarks/check-dams/compare.py \
  --remote-config .state/benchmarks/remote-upload/remote-config.json \
  --url http://39.106.98.67:28720 \
  --output .state/benchmarks/check-dams-wan-verified-20261108 \
  --warmups 1 --repeats 3 --timeout 1200