azure_ndmv4.yaml

# SuperBench Config
#
# Azure NDm A100 v4
# reference: https://docs.microsoft.com/en-us/azure/virtual-machines/ndm-a100-v4-series

version: v0.7
superbench:
  enable: null
  monitor:
    enable: true
    sample_duration: 1
    sample_interval: 10
  var:
    default_local_mode: &default_local_mode
      enable: true
      modes:
        - name: local
          proc_num: 8
          prefix: CUDA_VISIBLE_DEVICES={proc_rank}
          parallel: yes
    default_pytorch_mode: &default_pytorch_mode
      enable: true
      modes:
        - name: torch.distributed
          proc_num: 8
          node_num: 1
      frameworks:
        - pytorch
    common_model_config: &common_model_config
      duration: 0
      num_warmup: 64
      num_steps: 2048
      sample_count: 8192
      batch_size: 32
      precision:
        - float32
        - float16
      model_action:
        - train
      pin_memory: yes
  benchmarks:
    kernel-launch:
      <<: *default_local_mode
    gemm-flops:
      <<: *default_local_mode
    nccl-bw:default:
      enable: true
      modes:
        - name: local
          proc_num: 1
          parallel: no
      parameters:
        ngpus: 8
    nccl-bw:gdr-only:
      enable: true
      modes:
        - name: local
          proc_num: 1
          parallel: no
          env:
            NCCL_IB_PCI_RELAXED_ORDERING: '1'
            NCCL_NET_GDR_LEVEL: '5'
            NCCL_P2P_DISABLE: '1'
            NCCL_SHM_DISABLE: '1'
            NCCL_MIN_NCHANNELS: '16'
            NCCL_IB_DISABLE: '0'
      parameters:
        ngpus: 8
    ib-loopback:
      enable: true
      modes:
        - name: local
          proc_num: 4
          prefix: PROC_RANK={proc_rank} IB_DEVICES=0,2,4,6 NUMA_NODES=1,0,3,2
          parallel: yes
        - name: local
          proc_num: 4
          prefix: PROC_RANK={proc_rank} IB_DEVICES=1,3,5,7 NUMA_NODES=1,0,3,2
          parallel: yes
    cpu-memory-bw-latency:
      enable: false
      modes:
        - name: local
          proc_num: 1
          parallel: no
      parameters:
        tests:
          - bandwidth_matrix
          - latency_matrix
          - max_bandwidth
    mem-bw:
      enable: true
      modes:
        - name: local
          proc_num: 8
          prefix: CUDA_VISIBLE_DEVICES={proc_rank} numactl -N $(({proc_rank}/2))
          parallel: no
    disk-benchmark:
      enable: false
      modes:
        - name: local
          proc_num: 1
          parallel: no
      parameters:
        block_devices:
          - /dev/nvme0n1
          - /dev/nvme1n1
          - /dev/nvme2n1
          - /dev/nvme3n1
          - /dev/nvme4n1
          - /dev/nvme5n1
          - /dev/nvme6n1
          - /dev/nvme7n1
        seq_read_runtime: 60
        seq_write_runtime: 60
        seq_readwrite_runtime: 60
        rand_read_runtime: 60
        rand_write_runtime: 60
        rand_readwrite_runtime: 60
    gpu-copy-bw:correctness:
      enable: true
      modes:
        - name: local
          parallel: no
      parameters:
        mem_type:
          - htod
          - dtoh
          - dtod
        copy_type:
          - sm
          - dma
        size: 4096
        num_warm_up: 0
        num_loops: 1
        check_data: true
    gpu-copy-bw:perf:
      enable: true
      modes:
        - name: local
          parallel: no
      parameters:
        mem_type:
          - htod
          - dtoh
          - dtod
        copy_type:
          - sm
          - dma
    cudnn-function:
      <<: *default_local_mode
    cublas-function:
      <<: *default_local_mode
    matmul:
      <<: *default_local_mode
      frameworks:
        - pytorch
    sharding-matmul:
      <<: *default_pytorch_mode
    computation-communication-overlap:
      <<: *default_pytorch_mode
    ib-traffic:
      enable: false
      modes:
        - name: mpi
          proc_num: 8
      parameters:
        msg_size: 8388608
        ib_dev: mlx5_$LOCAL_RANK
        gpu_dev: $LOCAL_RANK
        numa_dev: $((LOCAL_RANK/2))
    gpcnet-network-test:
      enable: false
      modes:
        - name: mpi
          proc_num: 1
          mca:
            pml: ucx
            btl: ^uct
            btl_tcp_if_include: eth0
    gpcnet-network-load-test:
      enable: false
      modes:
        - name: mpi
          proc_num: 1
          mca:
            pml: ucx
            btl: ^uct
            btl_tcp_if_include: eth0
    tcp-connectivity:
      enable: false
      modes:
        - name: local
          parallel: no
      parameters:
        port: 22
    ort-inference:
      <<: *default_local_mode
    tensorrt-inference:
      <<: *default_local_mode
      parameters:
        pytorch_models:
          - resnet50
          - resnet101
          - resnet152
          - densenet169
          - densenet201
          - bert-base
          - bert-large
        seq_length: 224
        batch_size: 32
        precision: int8
    gpt_models:
      <<: *default_pytorch_mode
      models:
        - gpt2-small
        - gpt2-large
      parameters:
        <<: *common_model_config
        batch_size: 8
        seq_len: 224
    bert_models:
      <<: *default_pytorch_mode
      models:
        - bert-base
        - bert-large
      parameters:
        <<: *common_model_config
        seq_len: 224
    lstm_models:
      <<: *default_pytorch_mode
      models:
        - lstm
      parameters:
        <<: *common_model_config
        batch_size: 224
        input_size: 224
        hidden_size: 1000
        seq_len: 32
        pin_memory: no
    resnet_models:
      <<: *default_pytorch_mode
      models:
        - resnet50
        - resnet101
        - resnet152
      parameters:
        <<: *common_model_config
        batch_size: 192
        num_steps: 512
    densenet_models:
      <<: *default_pytorch_mode
      models:
        - densenet169
        - densenet201
      parameters:
        <<: *common_model_config
        pin_memory: no
    vgg_models:
      <<: *default_pytorch_mode
      models:
        - vgg11
        - vgg13
        - vgg16
        - vgg19
      parameters:
        <<: *common_model_config
        pin_memory: no