Skip to content

Latest commit

 

History

History
71 lines (62 loc) · 3.5 KB

followup_setup.md

File metadata and controls

71 lines (62 loc) · 3.5 KB

Followup Setup with Polyaxon v0.6.1

We have to configure polyaxon on the web UI at polyaxon 0.6.1, since the polyaxon configuration doesn't enable us to configure everything about polyaxon.

  • Get the pasword of root user
  • Configure custom Node Scheduling
  • Grant necessary GCP permission to the polyaxon's service account

Get the pasword of root user

root is the root user of polyaxon which is the original superuser by default. We can get the password the command below. We can login the polyaxon web UI as root with the password.

# The command derives from `helm status polyaxon`.
$ kubectl get secret --namespace polyaxon polyaxon-polyaxon-secret -o jsonpath="{.data.POLYAXON_ADMIN_PASSWORD}" | base64 --decode

Configure custom Node Scheduling

We can't configure custom node selectors of jobs, experiments and so on. Instead of polyaxon configuration, we have to modify the default scheduling on the web UI so that we assign polyaxon jobs and so on to appropriate node pools respectively. If we don't specify custom node scheduling, polyaxon jobs and so on will be assigned to any available resource. That is, polyaxon jobs potentially run on polyaxon core and builds node pools.

  • Builds Scheduling
    • URL: Builds Scheduling
    • key: NODE_SELECTORS:BUILD_JOBS
    • value: {"polyaxon": "builds"}
  • Jobs Scheduling
    • URL: Jobs Scheduling
    • key: NODE_SELECTORS:JOBS
    • value: {"polyaxon": "experiments"} or {"polyaxon": "experiments-preemptible"}
  • Experiments Scheduling
    • URL: Experiments Scheduling
    • key: NODE_SELECTORS:EXPERIMENTS
    • value: {"polyaxon": "experiments"} or {"polyaxon": "experiments-preemptible"}
  • Notebooks Scheduling
  • Tensorboards Scheduling

Cloud TPU

Unfortunately, we can specify only one Cloud TPU request resource in a polyaxon cluster with polyaxon v0. We can specify a Cloud TPU type and version under "Setting -> Hardware Accelerators" with K8S:TPU_TF_VERSION and K8S:TPU_RESOURCE_KEY respectively.:w

  • cloud-tpus.google.com/v2: TPU v2
  • cloud-tpus.google.com/preemptible-v2: Preemptible TPU v2
  • cloud-tpus.google.com/v3: TPU v3
  • cloud-tpus.google.com/preemptible-v3: Preemptible TPU v3

There is a list of Request a Cloud TPU in your Kubernetes Pod spec

Grant necessary GCP permission to the polyaxon's service account

The module provides an terraform output google_service_account_polyaxon to get the service account email of polyaxon. By registering the email to IAM in other GCP project, we can access to the resources. For instance, we probably need to access to BigQuery in other GCP project. We can also add google_service_account_polyaxon to your terraform project like output.tf of the basic example. Moreover, terraform output enables us to get the email.

$ terraform output
...
google_service_account_polyaxon = {
  "account_id" = "polyaxon"
  "display_name" = "polyaxon"
  "email" = "[email protected]"
}
...