Benchmarks four interleaved IO pipeline paths end-to-end on 80 local NVMe shards. All paths apply an aspect-ratio filter (1.0–2.0) to images.
| Path | Wall-clock time | Samples/sec | Input samples | Input size | Output rows | Output size |
|---|---|---|---|---|---|---|
| WDS → Parquet | 76.8 s | 88.8 | 6,818 | 9.9 GB | 57,713 rows | 3.84 GB |
| WDS → WDS | 75.4 s | 90.4 | 6,818 | 9.9 GB | 6,818 samples | 4.06 GB |
| PQ → Parquet | 15.7 s | 435.0 | 6,818 | 6.0 GB | 57,713 rows | 3.83 GB |
| #!/bin/bash | |
| # ═══════════════════════════════════════════════════════════════════════════════ | |
| # CONFIGURATION - Edit these values to match your deployment | |
| # ═══════════════════════════════════════════════════════════════════════════════ | |
| # Model names for each NIM service | |
| # Each service has an inference model and a pipeline model (for ensemble metrics) | |
| # Use "pipeline" if your NIM exposes generic pipeline metrics, or the ensemble name |
| ncountered Exception while running query | |
| Traceback (most recent call last): | |
| File "/home/vjawa/conda/envs/cudf_may_1/lib/python3.7/site-packages/xbb_tools-0.1-py3.7.egg/tools/utils.py", line 981, in run_dask_cudf_query | |
| results = query_func(client=client) | |
| File "/home/vjawa/conda/envs/cudf_may_1/lib/python3.7/site-packages/xbb_tools-0.1-py3.7.egg/tools/utils.py", line 308, in profiled | |
| result = func(*args, **kwargs) | |
| File "tpcx_bb_query_03.py", line 204, in main | |
| wcs_tstamp_min = get_wcs_minima(cli_args) | |
| File "tpcx_bb_query_03.py", line 39, in get_wcs_minima | |
| wcs_tstamp_min = wcs_df["tstamp"].min().compute() |
| <!DOCTYPE html> | |
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> |