Branch 3-0-2/bug fixes perf enhancements

.gitignore

@@ -63,3 +63,23 @@ env-fast
 
 # TLS certificates — local only, never commit (paths to certs are in .env)
 .certs/
+
+# Benchmark simulation output files
+sim_*.tsv
+sim_*.tsv.zst
+
+# Sweep run logs and results (local benchmark output)
+sweep_logs/
+sweep_flux_master.log
+results/
+
+# Test scripts and helpers not part of the benchmark suite
+test_s3dlio_gen_direct.py
+
+# Hydra runtime output (created in cwd when running workloads with hydra config)
+hydra_log/
+
+# Timestamped sweep run logs written to repo root by sweep_*.sh scripts
+sweep_unet3d_*.log
+sweep_dlrm_*.log
+sweep_flux_*.log

configs/dlio/workload/dlrm_b200.yaml

@@ -12,7 +12,7 @@ dataset:
   data_folder: data/dlrm/
   format: parquet
   num_files_train: 1024        # Number of training files to generate
-  num_samples_per_file: 4718592    # Samples per parquet file
+  num_samples_per_file: 1536000    # 250 RGs × 6144 → ~3.1 MiB footer (under s3-ultra 4 MiB limit)
   record_length_bytes: 761
   compression: none          # Options: snappy, gzip, lz4, zstd, none
 
@@ -627,12 +627,12 @@ dataset:
 reader:
   data_loader: pytorch
   batch_size: 12288
-  prefetch_size: 2  # Increase from default 2 for better I/O overlap
-  read_threads: 4   # Increase parallelism
+  prefetch_size: 0
+  read_threads: 0   # single-process, no IPC overhead; ThreadPoolExecutor handles I/O
   file_shuffle: seed
 
 train:
-  epochs: 1
+  epochs: 2
   computation_time: 0.000375
 
 metric:

configs/dlio/workload/dlrm_datagen.yaml

@@ -13,13 +13,14 @@ dataset:
   data_folder: data/dlrm/
   format: parquet
   num_files_train: 1024        # Number of training files to generate
-  num_samples_per_file: 4718592    # Samples per parquet file
+  num_samples_per_file: 1536000    # Samples per parquet file (250 RGs × 6144 → ~3.1 MiB footer, under s3-ultra 4 MiB limit)
   record_length_bytes: 761
   compression: none          # Options: snappy, gzip, lz4, zstd, none
 
   # Parquet-specific configuration
   parquet:
-    row_group_size: 8192
+    use_s3dlio_gen: true
+    row_group_size: 6144  # Match batch_size for optimal caching
     read_mode: row_group
 
     columns:

configs/dlio/workload/flux_datagen.yaml

@@ -17,6 +17,8 @@ dataset:
   record_length: 2164832
 
   parquet:
+    use_s3dlio_gen: true
+    row_group_size: 48
     # Parquet-specific field specifications
     columns:
       - name: t5_encodings

configs/dlio/workload/unet3d_b200.yaml

@@ -0,0 +1,40 @@
+model: 
+  name: unet3d
+  type: cnn
+  model_size: 499153191
+
+framework: pytorch
+
+workflow:
+  generate_data: False
+  train: True
+  checkpoint: False
+
+dataset: 
+  data_folder: data/unet3d/
+  format: npz
+  num_files_train: 7200        # ~984 GiB: 7200 × ~140 MiB avg file size
+  num_samples_per_file: 1
+  record_length_bytes: 146600628
+  record_length_bytes_stdev: 68341808
+  record_length_bytes_resize: 2097152
+
+reader: 
+  data_loader: pytorch
+  batch_size: 7
+  read_threads: 4
+  file_shuffle: seed
+  sample_shuffle: seed
+
+train:
+  epochs: 5
+  # B200 computation_time = H100 (0.323 s) ÷ 2 (B200 is ~2× faster than H100)
+  computation_time: 0.162
+
+checkpoint:
+  checkpoint_folder: checkpoints/unet3d
+  checkpoint_after_epoch: 5
+  epochs_between_checkpoints: 2
+
+metric:
+  au: 0.90