Initializing Azure instances is very slow

[suquark]

It takes me 14min to spin up a cluster with 2 cpu nodes.

The most time consuming part is installing pip packages, especially azure-cli. This may be addressed by releasing images with azure-cli pre-installed.

[infwinston]

+1. I had this slow initialization issue too. I might miss something but why is azure-cli needed to be install on remote VM?

[suquark]

Because ray-autoscaler is using it. For GCP and AWS, their CLIs are already installed.

[infwinston]

oh I see. It's used on the head node to further provision resources for worker nodes? Is it correct?

[suquark]

hmmm, it is mostly used by ray autoscaler for monitoring

[concretevitamin]

I tried revisiting this issue briefly. For a cpunode:

• Launching using Azure web console: about 1.5 min from "create" button to being able to SSH in. Same VM image, region. Only diff being using an existing resource group.
• Launching using sky launch (which means ray autoscaler, which means Azure python SDK): super slow, ~4-5min from create to SSH; total ~9 min (after installing runtime). Every step is slower than console.

I hacked the template by using the same resource group per region -- no speedup.

So the root cause seems to be Azure's python SDK being much slower than their console. We can take a deeper look.

Typical output

• ~4-5min from create to SSH
• ~4min to install runtime

I 08-25 08:58:46 cloud_vm_ray_backend.py:892] To view detailed progress: tail -n100 -f /Users/zongheng/sky_logs/sky-2022-08-25-08-58-44-664631/provision.log
I 08-25 08:58:46 cloud_vm_ray_backend.py:1096] Launching on Azure eastus ()
I 08-25 09:01:53 cloud_vm_ray_backend.py:1131] Retrying head node provisioning due to head fetching timeout.
I 08-25 09:03:40 log_utils.py:45] Head node is up.
I 08-25 09:07:41 cloud_vm_ray_backend.py:984] Successfully provisioned or found existing VM.

[romilbhardwaj]

So the root cause seems to be Azure's python SDK being much slower than their console. We can take a deeper look.

Might be good to verify this hypothesis by using their pure python SDK (without ray autoscaler) to provision a VM and measure time. Here's an example.

[github-actions[bot]]

This issue is stale because it has been open 120 days with no activity. Remove stale label or comment or this will be closed in 10 days.

[infwinston]

We should also keep this one open unless we are satisfied with the speed with Azure.

[github-actions[bot]]

This issue is stale because it has been open 120 days with no activity. Remove stale label or comment or this will be closed in 10 days.

[github-actions[bot]]

This issue was closed because it has been stalled for 10 days with no activity.

[WesleyYue]

Can this be re-opened? Still very slow today. For reference, a simple vllm setup takes 18 mins.

[WesleyYue]

Related https://github.com/skypilot-org/skypilot/issues/3695

[Michaelvll]

This issue should be mitigated by #3704. Closing for now.