Skip to content

Latest commit

 

History

History

bert

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
scripts
scripts
 
 
README.md
README.md
 
 
extract_paddle_logs.py
extract_paddle_logs.py
 
 

README.md

【DLPerf】PaddlePaddle BERT base测评

Overview

本次复现采用了PaddlePaddle官方仓库中的BERT，目的在于速度测评，同时根据测速结果给出1机、2机器、4机情况下的加速比，评判框架在分布式多机训练情况下的横向拓展能力。

目前，该测试覆盖了FP32、FP16混合精度，后续将持续维护，增加更多方式的测评。

Environment

系统

  • 系统：Ubuntu 16.04.4 LTS (GNU/Linux 4.4.0-116-generic x86_64)
  • 显卡：Tesla V100-SXM2-16GB x 8
  • 驱动：NVIDIA 440.33.01
  • CUDA：10.2
  • cuDNN：7.6.5
  • NCCL：2.7.3

框架

  • paddle 1.8.3.post107

Quick Start

项目代码

下载官方源码：

git clone https://github.com/PaddlePaddle/models/tree/release/1.8
cd models/PaddleNLP/pretrain_language_models/BERT

将本页面scripts路径下的脚本：make_pretrain_data.sh 放入BERT/data路径下，其余脚本全部放入：BERT/路径下

框架安装

python3 -m pip install paddlepaddle-gpu==1.8.3.post107 -i https://mirror.baidu.com/pypi/simple

NCCL

paddle的分布式训练底层依赖NCCL库，需要从NVIDIA-NCCL官网下载并安装和操作系统、CUDA版本适配的NCCL。 本次测试中安装2.7.3版本的NCCL：

sudo dpkg -i nccl-repo-ubuntu1604-2.7.3-ga-cuda10.2_1-1_amd64.deb
sudo apt update
sudo apt install libnccl2=2.7.3-1+cuda10.2 libnccl-dev=2.7.3-1+cuda10.2

数据集

本次BERT的预训练过程使用了paddle官方的示例数据集：demo_wiki_train.gz，由于数据集规模较小，我们在此基础上制作了demo_wiki_train_50.gz用于预训练。数据集制作过程如下：

cd models/PaddleNLP/pretrain_language_models/BERT/data
bash make_pretrain_data.sh

脚本将复制demo_wiki_train的内容，构造出一个50倍数据规模的训练集demo_wiki_train_50.gz。

Training

集群中有4台节点：

  • NODE1=10.11.0.2
  • NODE2=10.11.0.3
  • NODE3=10.11.0.4
  • NODE4=10.11.0.5

每个节点有8张显卡，这里设置batch size为32、64和96，分别在1机1卡～4机32卡的情况下进行了多组训练。

单机

models/PaddleNLP/pretrain_language_models/BERT目录下,执行脚本:

bash run_single_node.sh

对单机1卡、2卡、4卡、8卡分别做5次测试。单机多机脚本默认的batch size为32，可以通过参数指定，如指定batch size为64，bash run_single_node.sh 64，或96，bash run_single_node.sh 96。

2机16卡

2机、4机等多机情况下，需要在所有机器节点上相同路径准备同样的数据集、以完成分布式训练。

如2机：NODE1='10.11.0.2' NODE2='10.11.0.3' 的训练，需在两台机器上分别准备好数据集后，NODE1节点models/PaddleNLP/pretrain_language_models/BERT/目录下,执行脚本:

bash run_two_node.sh

NODE2节点models/PaddleNLP/pretrain_language_models/BERT/目录下，修改run_two_node.sh脚本中的CURRENT_NODE=$NODE2，再执行bash run_two_node.sh ，即可运行2机16卡的训练，同样默认测试5次。

4机32卡

流程同上，在4个机器节点上分别执行：

bash run_multi_node.sh

以运行4机32卡的训练，默认测试5次。

混合精度

运行FP16混合精度测试很简单，只需修改脚本参数或者运行时指定，例如通过如下命令：

bash run_multi_node.sh 64 fp16，即可运行batch size=64，FP16混合精度的测试

Result

吞吐率及加速比

执行以下命令，即可计算各种测试配置下的吞吐率及加速比：

python extract_paddle_logs.py --log_dir=logs/paddle/bert/bz64 --batch_size_per_device=64

输出：

logs/paddle/bert/bz64/4n8g/bert_b64_fp32_4.log {4: 2743.19}
logs/paddle/bert/bz64/4n8g/bert_b64_fp32_1.log {4: 2743.19, 1: 2699.39}
logs/paddle/bert/bz64/4n8g/bert_b64_fp32_2.log {4: 2743.19, 1: 2699.39, 2: 2745.97}
logs/paddle/bert/bz64/4n8g/bert_b64_fp32_6.log {4: 2743.19, 1: 2699.39, 2: 2745.97, 6: 2687.66}
logs/paddle/bert/bz64/4n8g/bert_b64_fp32_3.log {4: 2743.19, 1: 2699.39, 2: 2745.97, 6: 2687.66, 3: 2730.36}
logs/paddle/bert/bz64/4n8g/bert_b64_fp32_5.log {4: 2743.19, 1: 2699.39, 2: 2745.97, 6: 2687.66, 3: 2730.36, 5: 2745.92}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_4.log {4: 780.47}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_1.log {4: 780.47, 1: 756.94}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_2.log {4: 780.47, 1: 756.94, 2: 765.51}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_6.log {4: 780.47, 1: 756.94, 2: 765.51, 6: 744.27}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_3.log {4: 780.47, 1: 756.94, 2: 765.51, 6: 744.27, 3: 769.89}
logs/paddle/bert/bz64/1n8g/bert_b64_fp32_5.log {4: 780.47, 1: 756.94, 2: 765.51, 6: 744.27, 3: 769.89, 5: 737.23}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_4.log {4: 436.65}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_1.log {4: 436.65, 1: 463.53}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_2.log {4: 436.65, 1: 463.53, 2: 462.61}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_6.log {4: 436.65, 1: 463.53, 2: 462.61, 6: 441.4}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_3.log {4: 436.65, 1: 463.53, 2: 462.61, 6: 441.4, 3: 424.21}
logs/paddle/bert/bz64/1n4g/bert_b64_fp32_5.log {4: 436.65, 1: 463.53, 2: 462.61, 6: 441.4, 3: 424.21, 5: 442.09}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_4.log {4: 137.2}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_1.log {4: 137.2, 1: 137.06}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_2.log {4: 137.2, 1: 137.06, 2: 137.18}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_6.log {4: 137.2, 1: 137.06, 2: 137.18, 6: 137.35}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_3.log {4: 137.2, 1: 137.06, 2: 137.18, 6: 137.35, 3: 137.39}
logs/paddle/bert/bz64/1n1g/bert_b64_fp32_5.log {4: 137.2, 1: 137.06, 2: 137.18, 6: 137.35, 3: 137.39, 5: 137.59}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_4.log {4: 251.44}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_1.log {4: 251.44, 1: 252.99}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_2.log {4: 251.44, 1: 252.99, 2: 254.32}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_6.log {4: 251.44, 1: 252.99, 2: 254.32, 6: 252.04}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_3.log {4: 251.44, 1: 252.99, 2: 254.32, 6: 252.04, 3: 252.72}
logs/paddle/bert/bz64/1n2g/bert_b64_fp32_5.log {4: 251.44, 1: 252.99, 2: 254.32, 6: 252.04, 3: 252.72, 5: 252.7}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_4.log {4: 1418.26}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_1.log {4: 1418.26, 1: 1441.44}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_2.log {4: 1418.26, 1: 1441.44, 2: 1431.65}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_6.log {4: 1418.26, 1: 1441.44, 2: 1431.65, 6: 1389.89}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_3.log {4: 1418.26, 1: 1441.44, 2: 1431.65, 6: 1389.89, 3: 1447.72}
logs/paddle/bert/bz64/2n8g/bert_b64_fp32_5.log {4: 1418.26, 1: 1441.44, 2: 1431.65, 6: 1389.89, 3: 1447.72, 5: 1421.38}
{'bert': {'1n1g': {'average_speed': 137.29,
                   'batch_size_per_device': 64,
                   'median_speed': 137.27,
                   'speedup': 1.0},
          '1n2g': {'average_speed': 252.7,
                   'batch_size_per_device': 64,
                   'median_speed': 252.71,
                   'speedup': 1.84},
          '1n4g': {'average_speed': 445.08,
                   'batch_size_per_device': 64,
                   'median_speed': 441.74,
                   'speedup': 3.22},
          '1n8g': {'average_speed': 759.05,
                   'batch_size_per_device': 64,
                   'median_speed': 761.22,
                   'speedup': 5.55},
          '2n8g': {'average_speed': 1425.06,
                   'batch_size_per_device': 64,
                   'median_speed': 1426.52,
                   'speedup': 10.39},
          '4n8g': {'average_speed': 2725.42,
                   'batch_size_per_device': 64,
                   'median_speed': 2736.78,
                   'speedup': 19.94}}}
Saving result to ./result/bz64_result.json

计算规则

1.测速脚本

  • extract_paddle_logs.py

extract_paddle_logs.py根据官方在log中打印的速度，在120个iter中，排除前20iter，取后100个iter的速度做平均；

2.均值速度和中值速度

  • average_speed均值速度

  • median_speed中值速度

    每个batch size进行6次训练测试，记为一组，每一组取average_speed为均值速度，median_speed为中值速度。

3.加速比以中值速度计算

脚本和表格中的 加速比 是以单机单卡下的中值速度为基准进行计算的。例如:

单机单卡情况下速度为200(samples/s)，单机2卡速度为400，单机4卡速度为700，则加速比分别为：1.0、2.0、3.5

BERT-Base FP32

batch size=32 & without xla

node_num gpu_num samples/s speedup
1 1 132.64 1.00
1 2 228.12 1.72
1 4 406.02 3.06
1 8 615.12 4.64
2 16 1116.02 8.41
4 32 2073.6 15.63

batch size=64 & without xla

node_num gpu_num samples/s speedup
1 1 137.27 1.00
1 2 252.71 1.84
1 4 441.74 3.22
1 8 761.22 5.55
2 16 1426.52 10.39
4 32 2736.78 19.94

batch size=96 & without xla

node_num gpu_num samples/s speedup
1 1 136.97 1.00
1 2 258.73 1.89
1 4 490.38 3.58
1 8 868.6 6.34
2 16 1631.36 11.91
4 32 3167.68 23.13

BERT-Base FP16

batch size=64 & without xla

node_num gpu_num samples/s speedup
1 1 289.23 1
1 4 784.52 2.71
1 8 1298.96 4.49
2 16 1999.38 6.91
4 32 3406.36 11.78

batch size=64 & without xla & without dynamic_loss_scaling

without dynamic_loss_scaling即在测试中设置脚本(single_node_train.sh、multi_node_train.sh)参数：--use_dynamic_loss_scaling=false；开启动态loss scaling通常是为了解决fp16混合精度训练下的数据溢出问题，有助于模型收敛到正常精度，不过会略微影响训练速度。可以看见，关闭动态loss scaling后，单机单卡下训练速度由289.23samples/s提升至295.84samples/s，提升近2.29%

node_num gpu_num samples/s speedup
1 1 295.84 1
1 4 845.84 2.86
1 8 1471.45 4.97
2 16 2285.75 7.73
4 32 3801.84 12.85

batch size=128 & without xla

node_num gpu_num samples/s speedup
1 1 296.92 1
1 4 877.31 2.95
1 8 1538.25 5.18
2 16 2701.81 9.1
4 32 4922.16 16.58

batch size=160 & without xla

node_num gpu_num samples/s speedup
1 1 309.68 1
1 4 915.7 2.96
1 8 1666.54 5.38
2 16 2969.85 9.59
4 32 5452.35 17.61

完整日志