deploy/helm/health_checks/nccl_healthcheck/a3.yaml

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

job:
  base_name: "chs-hc"
  # guid: "xckd"  # Can specify a GUID if desired. Otherwise, a random GUID will be generated.
  # check_time: "1590303600"  # Will automatically be set if not given
health_check:
  name: "nccl"
  image:
    repo: "us-docker.pkg.dev/gce-ai-infra/health-check/nccl-healthcheck"
    tag: "a3-highgpu-8g_4.4.0"
    pull_policy: "Always"
  env:
    # A3 instance type is a3-highgpu-8g
    INSTANCE_TYPE: "a3-highgpu-8g"
    ENABLE_TWO_PASS_STRATEGY: "true"
    HEALTH_VALIDITY_HOURS: "24"
    DRY_RUN: "true"
    START_MESSAGE_SIZE: "2G"
    END_MESSAGE_SIZE: "8G"
    TEST_ITERATIONS: "3"
    NHOSTS: "2"
    nr: "8"
    # Specific to A3
    LD_LIBRARY_PATH: /usr/local/nvidia/lib64
    BANDWIDTH_THRESHOLD: "60"
    USE_TCPX: "true"
    USE_FASTRAK: "false"
    UNIX_CLIENT_PREFIX: "/run/tcpx"
    NCCL_DEBUG: "INFO"
    NCCL_DEBUG_SUBSYS: "INIT,NET"
    # Note A3 (a3-highgpu-8g) has no NCCL_LIB_DIR
    # Note A3 (a3-highgpu-8g) has no NCCL_FASTRAK_USE_SNAP
    # Note A3 (a3-highgpu-8g) has no NCCL_FASTRAK_ENABLE_CONTROL_CHANNEL
    # Note A3 (a3-highgpu-8g) has no NCCL_FASTRAK_NUM_FLOWS
  volumeMounts:
  - name: nvidia-install-dir-host
    mountPath: /usr/local/nvidia/lib64
  - name: tcpx-nccl-plugin-volume
    mountPath: /usr/local/tcpx
  - name: tcpd-socket
    mountPath: /run/tcpx
  - name: workload-terminated-volume
    mountPath: /usr/share/nemo
  - name: shared-memory
    mountPath: /dev/shm
  - name: varlog
    mountPath: /var/log

# Configuration specific to a3plus
volumes:
- name: nvidia-install-dir-host
  hostPath:
    path: /home/kubernetes/bin/nvidia/lib64
- name: tcpd-socket
  hostPath:
    path: /run/tcpx
- name: tcpx-nccl-plugin-volume
  emptyDir: {}
- name: workload-terminated-volume
  emptyDir: {}
- name: shared-memory
  emptyDir:
    medium: "Memory"
    sizeLimit: 200Gi
- name: varlog
  emptyDir: {}
- name: fluentbit-key
  secret:
    secretName: fluentbit-key
    optional: true
initContainers:
  nccl_plugin_installer:
    name: "nccl-plugin-installer"
    image: "us-docker.pkg.dev/gce-ai-infra/gpudirect-tcpx/nccl-plugin-gpudirecttcpx-dev:v3.1.7"
    imagePullPolicy: "Always"
    volumeMounts:
      name: "tcpx-nccl-plugin-volume"
      mountPath: "/usr/local/tcpx"
tcpd_daemon:
  image: "us-docker.pkg.dev/gce-ai-infra/gpudirect-tcpx/tcpgpudmarxd-dev:v2.0.12"
  imagePullPolicy: "Always"
  command:
  - "bash"
  args:
  - "-c"
  - |
    /tcpgpudmarxd/build/app/tcpgpudmarxd --gpu_nic_preset a3vm --gpu_shmem_type fd --setup_param "--verbose 128 2 0" --uds_path /run/tcpx &
    cleanup() {
      echo "Received SIGTERM or workload terminated, starting cleanup..."
      pkill -e "^"tcpgpudmarxd || true
      exit 0
    }
    trap cleanup SIGTERM
    while [ ! -e "/usr/share/nemo/workload_terminated" ]; do sleep 1; done
    cleanup
    sleep 10
  volumeMounts:
  - name: nvidia-install-dir-host
    mountPath: /usr/local/nvidia/lib64
  - name: tcpx-nccl-plugin-volume
    mountPath: /usr/local/tcpx
  - name: tcpd-socket
    mountPath: /run/tcpx
  - name: workload-terminated-volume
    mountPath: /usr/share/nemo