Dataset: Text (RedPajama) · Images · Image+Text (CC3M) · GPT-2 / ResNet-18 / CLIP · Tesla V100

LLM Data Pipeline Benchmark

AVG across 3 epochs · 3 loaders × 3 datasets × 3 batch sizes

WebDataset
Streaming (MosaicML)
Ray Data
Ray (pre-materialized shuffle) · bs=128 only
Text (RedPajama)
Images (w=2 parallel decode)
Image+Text (CC3M · CLIP)
Samples / sec ↑
GPU Utilisation %
Data Stall % ↓
GPU Memory (GB)
CPU Utilisation %
RAM Used (GB)
Elapsed (s) ↓ lower is better
Full Stats Table — Text
Metric wds-32wds-64wds-128 str-32str-64str-128 ray-32ray-64ray-128 ray-pms-128
samples/sec1,649.61,628.11,683.51,536.41,588.61,637.4450.0446.6444.01,678.1
GPU util %87.088.183.179.183.985.325.424.422.582.4
GPU mem GB2.823.113.752.823.113.752.823.113.752.13
CPU util %96.096.894.7138.2132.6133.669.469.265.8111.2
RAM used GB1.021.031.030.970.980.983.153.233.364.13
Data stall %27.827.328.420.527.735.572.572.573.81.32
Elapsed s8.08.17.88.68.38.129.429.629.87.87

w=2 workers · bs=32/64/128 · avg across 3 epochs · 100k images · parallel decode in DataLoader workers vs Ray object store

Samples / sec ↑
GPU Utilisation %
Data Stall % ↓
CPU Utilisation %
RAM Used (GB)
Elapsed (s) ↓
/dev/shm Avg Used (GB) — Ray object store pressure
/dev/shm Peak (GB)
Full Stats Table — Images · w=2 Parallel Decode · avg across 3 epochs
Metric wds-32wds-64wds-128 str-32str-64str-128 ray-32ray-64ray-128
samples/sec631.62726.42922.12937.211018.541084.05793.29896.381136.59
GPU util %20.3021.9525.6230.4231.6732.4423.4526.0734.82
GPU mem GB0.9501.1741.6230.9501.1741.6230.9501.1741.623
CPU util %1677.021259.671035.64116.93115.51112.601631.701502.581283.17
RAM used GB1.0821.1371.2001.1431.2031.3702.2932.3742.642
Data stall %52.6151.5055.741.450.810.6414.1221.4231.10
Elapsed s158.33137.78108.45106.7398.1792.34126.25111.7088.03
/dev/shm avg GB0.0100.0200.0400.0020.0040.0081.0301.0201.100
/dev/shm peak GB0.0190.0390.0770.0100.0100.0191.1631.1631.297
Samples / sec ↑
GPU Utilisation %
Data Stall % ↓
GPU Memory (GB)
CPU Utilisation %
RAM Used (GB)
Elapsed (s) ↓ lower is better
Full Stats Table — Image+Text (CC3M · CLIP)
Metric wds-32wds-64wds-128 str-32str-64str-128 ray-32ray-64ray-128
samples/sec22.9931.0137.3822.6031.2638.3022.7031.1837.47
GPU util %3.63.124.623.43.584.563.563.584.46
GPU mem GB3.093.223.493.093.223.493.093.223.50
CPU util %674.4908.01104.1670.4921.81101.1616.3834.6998.4
RAM used GB7.294.372.987.314.342.889.396.485.00
Data stall %0.650.861.040.440.730.913.715.206.43
Elapsed s434.9322.7267.5444.7321.4260.7440.6320.8266.8