Batch Jobs

Batch jobs is a feature to accommodate demands on non-emergency tasks with lower cost. Users can upload their tasks on a waiting queue managed by our platform, and we will run the task when the computing resources are available. Users can pay less than on-demand cost but get all computing resources runned at their need.

Prerequisites

Machine: An 8 GPU node. We would utilize PyTorchJob with one 8-GPU worker.

Image: Use DeepSpeed (0.14.2) Image in the Basic Image section.

Sample Task Setting

The mnist-distributed is used to train and test the performance of a simple CNN model on the FashionMNIST public dataset. We use this model to give an brief illustration on how this batch jobs feature works. The code is as following:

from __future__ import print_function

import argparse
import os

from tensorboardX import SummaryWriter
from torchvision import datasets, transforms
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

WORLD_SIZE = int(os.environ.get('WORLD_SIZE', 1))

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 20, 5, 1)
        self.conv2 = nn.Conv2d(20, 50, 5, 1)
        self.fc1 = nn.Linear(4*4*50, 500)
        self.fc2 = nn.Linear(500, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.max_pool2d(x, 2, 2)
        x = F.relu(self.conv2(x))
        x = F.max_pool2d(x, 2, 2)
        x = x.view(-1, 4*4*50)
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)
    
def train(args, model, device, train_loader, optimizer, epoch, writer):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tloss={:.4f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))
            niter = epoch * len(train_loader) + batch_idx
            writer.add_scalar('loss', loss.item(), niter)

def test(args, model, device, test_loader, writer, epoch):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
            pred = output.max(1, keepdim=True)[1] # get the index of the max log-probability
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)
    print('\naccuracy={:.4f}\n'.format(float(correct) / len(test_loader.dataset)))
    writer.add_scalar('accuracy', float(correct) / len(test_loader.dataset), epoch)

def should_distribute():
    return dist.is_available() and WORLD_SIZE > 1

def is_distributed():
    return dist.is_available() and dist.is_initialized()

def main():
    # Training settings
    parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
    parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                        help='input batch size for training (default: 64)')
    parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                        help='input batch size for testing (default: 1000)')
    parser.add_argument('--epochs', type=int, default=1, metavar='N',
                        help='number of epochs to train (default: 10)')
    parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                        help='learning rate (default: 0.01)')
    parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
                        help='SGD momentum (default: 0.5)')
    parser.add_argument('--no-cuda', action='store_true', default=False,
                        help='disables CUDA training')
    parser.add_argument('--seed', type=int, default=1, metavar='S',
                        help='random seed (default: 1)')
    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
                        help='how many batches to wait before logging training status')
    parser.add_argument('--save-model', action='store_true', default=False,
                        help='For Saving the current Model')
    parser.add_argument('--dir', default='logs', metavar='L',
                        help='directory where summary logs are stored')
    if dist.is_available():
        parser.add_argument('--backend', type=str, help='Distributed backend',
                            choices=[dist.Backend.GLOO, dist.Backend.NCCL, dist.Backend.MPI],
                            default=dist.Backend.GLOO)
    args = parser.parse_args()
    use_cuda = not args.no_cuda and torch.cuda.is_available()
    if use_cuda:
        print('Using CUDA')

    writer = SummaryWriter(args.dir)

    torch.manual_seed(args.seed)

    device = torch.device("cuda" if use_cuda else "cpu")

    if should_distribute():
        print('Using distributed PyTorch with {} backend'.format(args.backend))
        dist.init_process_group(backend=args.backend)

    kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
    train_loader = torch.utils.data.DataLoader(
        datasets.FashionMNIST('../data', train=True, download=True,
                       transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.batch_size, shuffle=True, **kwargs)
    test_loader = torch.utils.data.DataLoader(
        datasets.FashionMNIST('../data', train=False, transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.test_batch_size, shuffle=False, **kwargs)

    model = Net().to(device)

    if is_distributed():
        Distributor = nn.parallel.DistributedDataParallel if use_cuda \
            else nn.parallel.DistributedDataParallelCPU
        model = Distributor(model)

    optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)

    for epoch in range(1, args.epochs + 1):
        train(args, model, device, train_loader, optimizer, epoch, writer)
        test(args, model, device, test_loader, writer, epoch)

    if (args.save_model):
        torch.save(model.state_dict(),"mnist_cnn.pt")
        
if __name__ == '__main__':
    main()

We offer different types of storage (filesystem & highspeed) for storing code and datasets, taking filesystem as an example:

Users can create an instance to modify different types of storage, and this directory has been pre-configured with the aforementioned code file mnist.py.

Overall Process of Batch Jobs

1. Choose the appropriate machine configuration

2. Add job information. Currently, we use DeepSpeed image. We will be also supporting Colossal-AI distributed training engine in the near future.

3. We use PyTorchDDP framework.

4. Mount the shared storage and then input launch command.

pip install tensorboardX
python /root/highspeedstorage/test-mnist-highspeed/mnist.py --backend=nccl

5. After creation, the task will be running when resources are available.

Customized Dataset

The process mentioned above uses the FashionMNIST public dataset. Users can also configure their own datasets to shared storage. We enter the instance and modify the dataset directory in the code to the actual mounted directory.

train_loader = torch.utils.data.DataLoader(
        datasets.FashionMNIST('/root/filesystem/test-mnist/data/fashion', train=True, download=False,
                       transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.batch_size, shuffle=True, **kwargs)
    test_loader = torch.utils.data.DataLoader(
        datasets.FashionMNIST('/root/filesystem/test-mnist/data/fashion', train=False, transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=args.test_batch_size, shuffle=False, **kwargs)

Place the actual dataset in shared storage, where the public dataset is downloaded in advance.

cd /root/filesystem/test-mnist 

// git clone https://github.com/zalandoresearch/fashion-mnist.git
wget http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-images-idx3-ubyte.gz
wget http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/train-labels-idx1-ubyte.gz
wget http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-images-idx3-ubyte.gz
wget http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/t10k-labels-idx1-ubyte.gz

gunzip train-images-idx3-ubyte.gz
gunzip train-labels-idx1-ubyte.gz
gunzip t10k-images-idx3-ubyte.gz
gunzip t10k-labels-idx1-ubyte.gz

cp -r ~/filesystem/test-mnist/fashion-mnist/data/fashion/ ~/filesystem/test-mnist/data

Here is the new configuration.