前言

测试gpu前需要安装Anaconda、pytorch、tmux、nvitop。

单gpu

代码

import numpy as np
from tqdm import tqdm
'''
@Project ：gpu-test
@File    ：gpu_stress.py
@Author  ：xxx
@Date    ：2024/4/20 16:13
'''
import argparse
import time
import torch

if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='One gpu test')
    parser.add_argument('--size', type=int, required=False, default=10, help='矩阵的大小')
    parser.add_argument('--gpu', type=int, required=False, default=0, help='gpu的编号')
    parser.add_argument('--minu', type=int, required=False, default=1, help='运行时长：分')
    parser.add_argument('--verbose', type=int, required=False, default=10, help='进度条：运行时长的10%更新进度条')
    args = parser.parse_args()
    # 设置要运行的时长（分钟）
    duration_mins = args.minu
    verbose = args.verbose
    torch.cuda.set_device(args.gpu)
    size = args.size

    # 测试 GPU 计算耗时

    A = torch.ones(size, size).to('cuda')
    B = torch.ones(size, size).to('cuda')

    start_time = time.time()
    current_time = start_time
    #创建一个进度条
    ratio = verbose/100  #计算当前进度
    with tqdm(total=100) as pbar:
        pbar.set_description("Processing")
        while (current_time - start_time) / 60 < duration_mins:
            C = torch.inner(A, B)
            current_time = time.time()
            if np.round(((current_time - start_time)/60)/duration_mins,1) ==  np.round(ratio,1):
                pbar.update(verbose)#更新进度条
                ratio += verbose/100

    total_duration_seconds = current_time - start_time
    total_duration_hours = total_duration_seconds / 3600
    print(f"GPU RUN TIME: {total_duration_hours:.2f} 小时")

运行命令

python gpu.py --gpu 0 --size 44000 --minu 6000 
python gpu.py --gpu 1 --size 44000 --minu 6000 

参数说明：

• gpu：显卡编号
• size：矩阵大小，如果双卡32G显存，可以设置参数为44000，其他情况慢慢试
• minu：运行时长，单位-分钟

多gpu并行

代码

#!/usr/bin/env python
# -*- coding: utf-8 -*-
'''
@Project ：gpu-test
@File    ：gpus_stress.py
@Author  ：xxx
@Date    ：2024/4/20 16:13
'''
import argparse
import time

import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
from tqdm import tqdm


def train_model(duration_mins, num_gpus,verbose):
    duration_secs = duration_mins * 60
    size = args.size
    data = torch.ones(size, size).to('cuda')

    class MyModel(nn.Module):
        def __init__(self):
            super(MyModel, self).__init__()
            self.fc1 = nn.Linear(size, size)

        def forward(self, x):
            return self.fc1(x)

    model = MyModel().to('cuda')
    model = nn.DataParallel(model, device_ids=list(range(num_gpus)))
    optimizer = optim.SGD(model.parameters(), lr=0.001)
    criterion = nn.MSELoss()

    start_time = time.time()
    current_time = start_time
    # 创建一个进度条
    ratio = verbose / 100  # 计算当前进度
    with tqdm(total=100) as pbar:
        pbar.set_description("Processing")
        while (current_time - start_time) < duration_secs:
            optimizer.zero_grad()
            outputs = model(data)
            loss = criterion(outputs, data)
            loss.backward()
            optimizer.step()
            current_time = time.time()
            if np.round(((current_time - start_time) / 60) / duration_mins, 1) == np.round(ratio, 1):
                pbar.update(verbose)  # 更新进度条
                ratio += verbose / 100

    total_duration_hours = (current_time - start_time) / 3600
    print(f"GPU RUN TIME: {total_duration_hours:.2f} hours")


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Multi-GPU Training')
    parser.add_argument('--minu', type=int, default=1, help='Duration in minutes')
    parser.add_argument('--gpus', type=int, default=1, help='Number of GPUs to use')
    parser.add_argument('--size', type=int, required=False, default=1, help='size')
    parser.add_argument('--verbose', type=int, required=False, default=10, help='进度条：运行时长的10%更新进度条')
    args = parser.parse_args()

    train_model(args.minu, args.gpus, args.verbose)

运行命令

32g显存的显卡执行以下命令

python gpus.py --gpus 2 --size 30000 --minu 6000

参数说明：

• gpus：可用显卡数量
• size：矩阵大小，如果双卡32G显存，可以设置参数为30000，其他情况慢慢试
• minu：运行时长，单位-分钟