Container Service for Kubernetes (ACK) supports topology-aware GPU scheduling based on the scheduling framework. This feature selects a combination of GPUs from GPU-accelerated nodes to achieve optimal GPU acceleration for training jobs. This topic describes how to use topology-aware GPU scheduling to achieve optimal GPU acceleration for PyTorch distributed jobs.

Prerequisites

  • Create a professional managed Kubernetes cluster.
  • Arena is installed.
  • Install ack-ai-installer.
  • The following table lists the required components and versions.
    Component Version
    Kubernetes V1.18.8 and later
    Helm V3.0 and later
    Nvidia V418.87.01 and later
    NVIDIA Collective Communications Library (NCCL) 2.7+
    Docker 19.03.5
    Operating system CentOS 7.6, CentOS 7.7, Ubuntu 16.04 and 18.04, and Alibaba Cloud Linux 2.
    GPU V100

Limits

  • Topology-aware GPU scheduling is applicable to only Message Passing Interface (MPI) jobs that are trained by using a distributed framework.
  • The resources that are requested by pods must meet specific requirements before the pods can be created to submit and start jobs. Otherwise, the requests remain pending for resources.

Procedure

Configure nodes

Run the following command to set the node label and explicitly enable topology-aware GPU scheduling for nodes:
kubectl label node <Your Node Name> ack.node.gpu.schedule=topology
Note After topology-aware GPU scheduling is enabled on nodes, common GPU scheduling is no longer supported. You can run the following command to change the label and resume common GPU scheduling:
kubectl label node <Your Node Name> ack.node.gpu.schedule=default --overwrite

Submit a job

Submit a Message Passing Interface (MPI) job and set --gputopology to true.

arena submit mpi --gputopology=true ***

Example 1: Train VGG16

Note In this topic, two servers are deployed in the test cluster. Each server has eight V100 GPUs.

Use topology-aware GPU scheduling to train VGG16

  1. Run the following command to submit a job to the cluster:
    arena submit mpi \
      --name=pytorch-topo-4-vgg16 \
      --gpus=1 \
      --workers=4 \
      --gputopology=true \
      --image=registry.cn-hangzhou.aliyuncs.com/kubernetes-image-hub/pytorch-benchmark:torch1.6.0-py3.7-cuda10.1 \
      "mpirun --allow-run-as-root -np "4" -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x NCCL_SOCKET_IFNAME=eth0 -x LD_LIBRARY_PATH -x PATH --mca pml ob1 --mca btl_tcp_if_include eth0 --mca oob_tcp_if_include eth0 --mca orte_keep_fqdn_hostnames t --mca btl ^openib python /examples/pytorch_synthetic_benchmark.py --model=vgg16 --batch-size=64"
  2. Run the following command to query the state of the job:
    arena get pytorch-topo-4-vgg16 --type mpijob

    Expected output:

    Name:      pytorch-topo-4-vgg16
    Status:    RUNNING
    Namespace: default
    Priority:  N/A
    Trainer:   MPIJOB
    Duration:  11s
    
    Instances:
      NAME                                 STATUS   AGE  IS_CHIEF  GPU(Requested)  NODE
      ----                                 ------   ---  --------  --------------  ----
      pytorch-topo-4-vgg16-launcher-mnjzr  Running  11s  true      0               cn-shanghai.192.168.16.173
      pytorch-topo-4-vgg16-worker-0        Running  11s  false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-vgg16-worker-1        Running  11s  false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-vgg16-worker-2        Running  11s  false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-vgg16-worker-3        Running  11s  false     1               cn-shanghai.192.168.16.173
  3. Run the following command to print the job log:
    arena logs -f pytorch-topo-4-vgg16

    Expected output:

    Model: vgg16
    Batch size: 64
    Number of GPUs: 4
    Running warmup...
    Running benchmark...
    Iter #0: 205.5 img/sec per GPU
    Iter #1: 205.2 img/sec per GPU
    Iter #2: 205.1 img/sec per GPU
    Iter #3: 205.5 img/sec per GPU
    Iter #4: 205.1 img/sec per GPU
    Iter #5: 205.1 img/sec per GPU
    Iter #6: 205.3 img/sec per GPU
    Iter #7: 204.3 img/sec per GPU
    Iter #8: 205.0 img/sec per GPU
    Iter #9: 204.9 img/sec per GPU
    Img/sec per GPU: 205.1 +-0.6
    Total img/sec on 4 GPU(s): 820.5 +-2.5

Use common GPU scheduling to train VGG16

  1. Run the following command to submit a job to the cluster:
    arena submit mpi \
      --name=pytorch-4-vgg16 \
      --gpus=1 \
      --workers=4 \
      --image=registry.cn-hangzhou.aliyuncs.com/kubernetes-image-hub/pytorch-benchmark:torch1.6.0-py3.7-cuda10.1 \
      "mpirun --allow-run-as-root -np "4" -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x NCCL_SOCKET_IFNAME=eth0 -x LD_LIBRARY_PATH -x PATH --mca pml ob1 --mca btl_tcp_if_include eth0 --mca oob_tcp_if_include eth0 --mca orte_keep_fqdn_hostnames t --mca btl ^openib python /examples/pytorch_synthetic_benchmark.py --model=vgg16 --batch-size=64"
  2. Run the following command to query the state of the job:
    arena get pytorch-4-vgg16 --type mpijob

    Expected output:

    Name:      pytorch-4-vgg16
    Status:    RUNNING
    Namespace: default
    Priority:  N/A
    Trainer:   MPIJOB
    Duration:  10s
    
    Instances:
      NAME                            STATUS   AGE  IS_CHIEF  GPU(Requested)  NODE
      ----                            ------   ---  --------  --------------  ----
      pytorch-4-vgg16-launcher-qhnxl  Running  10s  true      0               cn-shanghai.192.168.16.173
      pytorch-4-vgg16-worker-0        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-vgg16-worker-1        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-vgg16-worker-2        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-vgg16-worker-3        Running  10s  false     1               cn-shanghai.192.168.16.173
  3. Run the following command to print the job log:
    arena logs -f pytorch-4-vgg16

    Expected output:

    Model: vgg16
    Batch size: 64
    Number of GPUs: 4
    Running warmup...
    Running benchmark...
    Iter #0: 113.1 img/sec per GPU
    Iter #1: 109.5 img/sec per GPU
    Iter #2: 106.5 img/sec per GPU
    Iter #3: 108.5 img/sec per GPU
    Iter #4: 108.1 img/sec per GPU
    Iter #5: 111.2 img/sec per GPU
    Iter #6: 110.7 img/sec per GPU
    Iter #7: 109.8 img/sec per GPU
    Iter #8: 102.8 img/sec per GPU
    Iter #9: 107.9 img/sec per GPU
    Img/sec per GPU: 108.8 +-5.3
    Total img/sec on 4 GPU(s): 435.2 +-21.1

Example 2: Train ResNet50

Use topology-aware GPU scheduling to train ResNet50

  1. Run the following command to submit a job to the cluster:
    arena submit mpi \
      --name=pytorch-topo-4-resnet50 \
      --gpus=1 \
      --workers=4 \
      --gputopology=true \
      --image=registry.cn-hangzhou.aliyuncs.com/kubernetes-image-hub/pytorch-benchmark:torch1.6.0-py3.7-cuda10.1 \
      "mpirun --allow-run-as-root -np "4" -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x NCCL_SOCKET_IFNAME=eth0 -x LD_LIBRARY_PATH -x PATH --mca pml ob1 --mca btl_tcp_if_include eth0 --mca oob_tcp_if_include eth0 --mca orte_keep_fqdn_hostnames t --mca btl ^openib python /examples/pytorch_synthetic_benchmark.py --model=resnet50 --batch-size=64"
  2. Run the following command to query the state of the job:
    arena get pytorch-topo-4-resnet50 --type mpijob

    Expected output:

    Name:      pytorch-topo-4-resnet50
    Status:    RUNNING
    Namespace: default
    Priority:  N/A
    Trainer:   MPIJOB
    Duration:  8s
    
    Instances:
      NAME                                    STATUS   AGE  IS_CHIEF  GPU(Requested)  NODE
      ----                                    ------   ---  --------  --------------  ----
      pytorch-topo-4-resnet50-launcher-x7r2n  Running  8s   true      0               cn-shanghai.192.168.16.173
      pytorch-topo-4-resnet50-worker-0        Running  8s   false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-resnet50-worker-1        Running  8s   false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-resnet50-worker-2        Running  8s   false     1               cn-shanghai.192.168.16.173
      pytorch-topo-4-resnet50-worker-3        Running  8s   false     1               cn-shanghai.192.168.16.173
  3. Run the following command to print the job log:
    arena logs -f pytorch-topo-4-resnet50

    Expected output:

    Model: resnet50
    Batch size: 64
    Number of GPUs: 4
    Running warmup...
    Running benchmark...
    Iter #0: 331.0 img/sec per GPU
    Iter #1: 330.6 img/sec per GPU
    Iter #2: 330.9 img/sec per GPU
    Iter #3: 330.4 img/sec per GPU
    Iter #4: 330.7 img/sec per GPU
    Iter #5: 330.8 img/sec per GPU
    Iter #6: 329.9 img/sec per GPU
    Iter #7: 330.5 img/sec per GPU
    Iter #8: 330.4 img/sec per GPU
    Iter #9: 329.7 img/sec per GPU
    Img/sec per GPU: 330.5 +-0.8
    Total img/sec on 4 GPU(s): 1321.9 +-3.2

Use common GPU scheduling to train ResNet50

  1. Run the following command to submit a job to the cluster:
    arena submit mpi \
      --name=pytorch-4-resnet50 \
      --gpus=1 \
      --workers=4 \
      --image=registry.cn-hangzhou.aliyuncs.com/kubernetes-image-hub/pytorch-benchmark:torch1.6.0-py3.7-cuda10.1 \
      "mpirun --allow-run-as-root -np "4" -bind-to none -map-by slot -x NCCL_DEBUG=INFO -x NCCL_SOCKET_IFNAME=eth0 -x LD_LIBRARY_PATH -x PATH --mca pml ob1 --mca btl_tcp_if_include eth0 --mca oob_tcp_if_include eth0 --mca orte_keep_fqdn_hostnames t --mca btl ^openib python /examples/pytorch_synthetic_benchmark.py --model=resnet50 --batch-size=64"
  2. Run the following command to query the state of the job:
    arena get pytorch-4-resnet50 --type mpijob

    Expected output:

    Name:      pytorch-4-resnet50
    Status:    RUNNING
    Namespace: default
    Priority:  N/A
    Trainer:   MPIJOB
    Duration:  10s
    
    Instances:
      NAME                               STATUS   AGE  IS_CHIEF  GPU(Requested)  NODE
      ----                               ------   ---  --------  --------------  ----
      pytorch-4-resnet50-launcher-qw5k6  Running  10s  true      0               cn-shanghai.192.168.16.173
      pytorch-4-resnet50-worker-0        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-resnet50-worker-1        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-resnet50-worker-2        Running  10s  false     1               cn-shanghai.192.168.16.173
      pytorch-4-resnet50-worker-3        Running  10s  false     1               cn-shanghai.192.168.16.173
  3. Run the following command to print the job log:
    arena logs -f pytorch-4-resnet50

    Expected output:

    Model: resnet50
    Batch size: 64
    Number of GPUs: 4
    Running warmup...
    Running benchmark...
    Iter #0: 313.1 img/sec per GPU
    Iter #1: 312.8 img/sec per GPU
    Iter #2: 313.0 img/sec per GPU
    Iter #3: 312.2 img/sec per GPU
    Iter #4: 313.7 img/sec per GPU
    Iter #5: 313.2 img/sec per GPU
    Iter #6: 313.6 img/sec per GPU
    Iter #7: 313.0 img/sec per GPU
    Iter #8: 311.3 img/sec per GPU
    Iter #9: 313.6 img/sec per GPU
    Img/sec per GPU: 313.0 +-1.3
    Total img/sec on 4 GPU(s): 1251.8 +-5.3

Performance comparison

The following figure shows the performance comparison between topology-aware GPU scheduling and common GPU scheduling based on the preceding examples. gpu32
The figure shows that after topology-aware GPU scheduling is activated, the PyTorch distributed jobs are significantly accelerated.
Note The improvement that is achieved by topology-aware GPU scheduling varies based on the models that you use and the cluster environment. You can evaluate the performance of your models based on the preceding examples.