[Performance] Speed up Azure A10 instance creation (skypilot-org#4205)

* Use date instead of timestamp in skypilot image names * Speed up Azure A10 VM creation * disable nouveau and use smaller instance * address comments * address comments * add todo
cblmemo · Nov 7, 2024 · 83fade9 · 83fade9
1 parent 30136b7
commit 83fade9
Show file tree

Hide file tree

Showing 12 changed files with 69 additions and 59 deletions.
diff --git a/sky/backends/cloud_vm_ray_backend.py b/sky/backends/cloud_vm_ray_backend.py
@@ -1950,17 +1950,8 @@ def provision_with_retries(
 
         failover_history: List[Exception] = list()
 
-        style = colorama.Style
-        fore = colorama.Fore
         # Retrying launchable resources.
         while True:
-            if (isinstance(to_provision.cloud, clouds.Azure) and
-                    to_provision.accelerators is not None and
-                    'A10' in to_provision.accelerators and prev_handle is None):
-                logger.warning(f'{style.BRIGHT}{fore.YELLOW}Trying to launch '
-                               'an A10 cluster on Azure. This may take ~20 '
-                               'minutes due to driver installation.'
-                               f'{style.RESET_ALL}')
             try:
                 # Recheck cluster name as the 'except:' block below may
                 # change the cloud assignment.

diff --git a/sky/clouds/azure.py b/sky/clouds/azure.py
@@ -44,6 +44,8 @@
 _DEFAULT_V1_IMAGE_ID = 'skypilot:custom-gpu-ubuntu-v1'
 _DEFAULT_GPU_K80_IMAGE_ID = 'skypilot:k80-ubuntu-2004'
 _FALLBACK_IMAGE_ID = 'skypilot:gpu-ubuntu-2204'
+# This is used by Azure GPU VMs that use grid drivers (e.g. A10).
+_DEFAULT_GPU_GRID_IMAGE_ID = 'skypilot:custom-gpu-ubuntu-v2-grid'
 
 _COMMUNITY_IMAGE_PREFIX = '/CommunityGalleries'
 
@@ -220,6 +222,8 @@ def _get_default_image_tag(self, gen_version, instance_type) -> str:
             acc_name = list(acc.keys())[0]
             if acc_name == 'K80':
                 return _DEFAULT_GPU_K80_IMAGE_ID
+            if acc_name == 'A10':
+                return _DEFAULT_GPU_GRID_IMAGE_ID
         # About Gen V1 vs V2:
         # In Azure, all instances with K80 (Standard_NC series), some
         # instances with M60 (Standard_NV series) and some cpu instances
@@ -350,10 +354,6 @@ def make_deploy_resources_variables(
                 'image_version': version,
             }
 
-        # Setup the A10 nvidia driver.
-        need_nvidia_driver_extension = (acc_dict is not None and
-                                        'A10' in acc_dict)
-
         # Determine resource group for deploying the instance.
         resource_group_name = skypilot_config.get_nested(
             ('azure', 'resource_group_vm'), None)
@@ -413,7 +413,6 @@ def _failover_disk_tier() -> Optional[resources_utils.DiskTier]:
             # Azure does not support specific zones.
             'zones': None,
             **image_config,
-            'need_nvidia_driver_extension': need_nvidia_driver_extension,
             'disk_tier': Azure._get_disk_type(disk_tier),
             'cloud_init_setup_commands': cloud_init_setup_commands,
             'azure_subscription_id': self.get_project_id(dryrun),

diff --git a/sky/clouds/service_catalog/images/README.md b/sky/clouds/service_catalog/images/README.md
@@ -59,9 +59,10 @@ export SECRET=xxxxxx  # Update this
 ```
 2. Build and copy images for all regions for GPU (gen 1 & 2) and CPU (gen 2 only).
 ```bash
-export TYPE=gpu  # Update this
-export VM_GENERATION=1  # Update this
-packer build --var vm_generation=${VM_GENERATION} --var client_secret=${SECRET} skypilot-azure-${TYPE}-ubuntu.pkr.hcl
+packer build --var vm_generation=2 --var client_secret=${SECRET} skypilot-azure-cpu-ubuntu.pkr.hcl
+packer build --var vm_generation=2 --var client_secret=${SECRET} skypilot-azure-gpu-ubuntu.pkr.hcl
+packer build --var vm_generation=1 --var client_secret=${SECRET} skypilot-azure-gpu-ubuntu.pkr.hcl
+packer build  --var vm_generation=2 --var client_secret=${SECRET} --var use_grid_driver=true skypilot-azure-gpu-ubuntu.pkr.hcl
 ```
 
 ## Test Images

diff --git a/sky/clouds/service_catalog/images/provisioners/cuda-azure-grid.sh b/sky/clouds/service_catalog/images/provisioners/cuda-azure-grid.sh
@@ -0,0 +1,33 @@
+#!/bin/bash
+
+sudo apt update
+sudo apt install -y build-essential
+
+echo "Installing GRID driver..."
+GRID_DRIVER_URL="https://download.microsoft.com/download/8/d/a/8da4fb8e-3a9b-4e6a-bc9a-72ff64d7a13c/NVIDIA-Linux-x86_64-535.161.08-grid-azure.run"
+GRID_DRIVER_FILE="NVIDIA-Linux-x86_64-535.161.08-grid-azure.run"
+
+wget -nv $GRID_DRIVER_URL -O $GRID_DRIVER_FILE
+sudo chmod +x $GRID_DRIVER_FILE
+sudo sh $GRID_DRIVER_FILE --silent --disable-nouveau
+
+echo "Set vGPU Licensing Daemon config..."
+sudo cp /etc/nvidia/gridd.conf.template /etc/nvidia/gridd.conf
+sudo sed -i '/^FeatureType=0/s/^/# /' /etc/nvidia/gridd.conf
+echo "IgnoreSP=FALSE" | sudo tee -a /etc/nvidia/gridd.conf
+echo "EnableUI=FALSE" | sudo tee -a /etc/nvidia/gridd.conf
+
+echo "Installing CUDA toolkit..."
+CUDA_TOOLKIT_URL="https://developer.download.nvidia.com/compute/cuda/12.2.0/local_installers/cuda_12.2.0_535.54.03_linux.run"
+CUDA_TOOLKIT_FILE="cuda_12.2.0_535.54.03_linux.run"
+wget -nv $CUDA_TOOLKIT_URL -O $CUDA_TOOLKIT_FILE
+sudo sh $CUDA_TOOLKIT_FILE --silent --toolkit --override
+
+# Set environment variables
+echo 'export PATH=$PATH:/usr/local/cuda-12.2/bin' >> ~/.bashrc
+echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-12.2/lib64' >> ~/.bashrc
+source ~/.bashrc
+
+# Verify installations
+rm -f NVIDIA-Linux-x86_64-535.161.08-grid-azure.run cuda_12.2.0_535.54.03_linux.run
+nvidia-smi
diff --git a/sky/clouds/service_catalog/images/skypilot-aws-cpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-aws-cpu-ubuntu.pkr.hcl
@@ -4,11 +4,11 @@ variable "region" {
 }
 
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
 }
 
 source "amazon-ebs" "cpu-ubuntu" {
-  ami_name      = "skypilot-aws-cpu-ubuntu-${local.timestamp}"
+  ami_name      = "skypilot-aws-cpu-ubuntu-${local.date}"
   instance_type = "t2.micro"
   region        = var.region
   ssh_username  = "ubuntu"

diff --git a/sky/clouds/service_catalog/images/skypilot-aws-gpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-aws-gpu-ubuntu.pkr.hcl
@@ -4,11 +4,11 @@ variable "region" {
 }
 
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
 }
 
 source "amazon-ebs" "gpu-ubuntu" {
-  ami_name      = "skypilot-aws-gpu-ubuntu-${local.timestamp}"
+  ami_name      = "skypilot-aws-gpu-ubuntu-${local.date}"
   instance_type = "g6.xlarge"
   region        = var.region
   ssh_username  = "ubuntu"

diff --git a/sky/clouds/service_catalog/images/skypilot-azure-cpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-azure-cpu-ubuntu.pkr.hcl
@@ -9,13 +9,14 @@ variable "vm_generation" {
 }
 
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
   version   = formatdate("YY.MM.DD", timestamp())
 }
 
 source "azure-arm" "cpu-ubuntu" {
   managed_image_resource_group_name = "skypilot-images"
-  managed_image_name                = "skypilot-azure-cpu-ubuntu-${local.timestamp}"
+  // TODO(yika): these fields may not be required as we use community images below instead. We need to double-check if these can be removed.
+  managed_image_name                = "skypilot-azure-cpu-ubuntu-${local.date}"
 
   subscription_id = "59d8c23c-7ef5-42c7-b2f3-a919ad8026a7"
   tenant_id       = "7c81f068-46f8-4b26-9a46-2fbec2287e3d"

diff --git a/sky/clouds/service_catalog/images/skypilot-azure-gpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-azure-gpu-ubuntu.pkr.hcl
@@ -8,14 +8,20 @@ variable "vm_generation" {
   description = "Azure's VM generation, currently support 1 or 2"
 }
 
+variable "use_grid_driver" {
+  type        = bool
+  default     = false
+  description = "Whether to use the Azure GRID driver. Currently only A10 GPU VMs need this."
+}
+
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
   version   = formatdate("YY.MM.DD", timestamp())
 }
 
 source "azure-arm" "gpu-ubuntu" {
   managed_image_resource_group_name = "skypilot-images"
-  managed_image_name                = "skypilot-azure-gpu-ubuntu-${local.timestamp}"
+  managed_image_name                = "skypilot-azure-gpu-ubuntu-${local.date}"
 
   subscription_id = "59d8c23c-7ef5-42c7-b2f3-a919ad8026a7"
   tenant_id       = "7c81f068-46f8-4b26-9a46-2fbec2287e3d"
@@ -26,8 +32,8 @@ source "azure-arm" "gpu-ubuntu" {
   image_publisher = "Canonical"
   image_offer     = "0001-com-ubuntu-server-jammy"
   image_sku       = var.vm_generation == 1 ? "22_04-lts" : "22_04-lts-gen2"
-  location        = var.vm_generation == 1 ? "eastus" : "centralus"
-  vm_size         = var.vm_generation == 1 ? "Standard_NC4as_T4_v3" : "Standard_NC24ads_A100_v4"
+  location        = var.use_grid_driver || var.vm_generation == 1 ? "eastus" : "centralus"
+  vm_size         = var.use_grid_driver ? "Standard_NV12ads_A10_v5" : (var.vm_generation == 1 ? "Standard_NC4as_T4_v3" : "Standard_NC24ads_A100_v4")
   ssh_username    = "azureuser"
   azure_tags = {
     Created_by = "packer"
@@ -37,8 +43,8 @@ source "azure-arm" "gpu-ubuntu" {
   shared_image_gallery_destination {
     subscription   = "59d8c23c-7ef5-42c7-b2f3-a919ad8026a7"
     resource_group = "skypilot-images"
-    gallery_name   = var.vm_generation == 1 ? "skypilot_images": "skypilot_image_gallery"
-    image_name     = "skypilot-gpu-gen${var.vm_generation}"
+    gallery_name   = var.use_grid_driver || var.vm_generation == 1 ? "skypilot_images" : "skypilot_image_gallery"
+    image_name     = var.use_grid_driver ? "skypilot-gpu-gen2-grid" : "skypilot-gpu-gen${var.vm_generation}"
     image_version  = "${local.version}"
     replication_regions = [
       "centralus",
@@ -61,7 +67,7 @@ build {
     script = "./provisioners/docker.sh"
   }
   provisioner "shell" {
-    script = "./provisioners/cuda.sh"
+    script = var.use_grid_driver ? "./provisioners/cuda-azure-grid.sh" : "./provisioners/cuda.sh"
   }
   provisioner "shell" {
     script = "./provisioners/nvidia-container-toolkit.sh"

diff --git a/sky/clouds/service_catalog/images/skypilot-gcp-cpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-gcp-cpu-ubuntu.pkr.hcl
@@ -1,11 +1,11 @@
 
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
 }
 
 source "googlecompute" "cpu-ubuntu" {
   project_id          = "sky-dev-465"
-  image_name          = "skypilot-gcp-cpu-ubuntu-${local.timestamp}"
+  image_name          = "skypilot-gcp-cpu-ubuntu-${local.date}"
   source_image_family = "ubuntu-2204-lts"
   zone                = "us-west1-a"
   image_description   = "SkyPilot custom image for launching GCP CPU instances."

diff --git a/sky/clouds/service_catalog/images/skypilot-gcp-gpu-ubuntu.pkr.hcl b/sky/clouds/service_catalog/images/skypilot-gcp-gpu-ubuntu.pkr.hcl
@@ -4,11 +4,11 @@ variable "zone" {
 }
 
 locals {
-  timestamp = regex_replace(timestamp(), "[- TZ:]", "")
+  date = formatdate("YYMMDD", timestamp())
 }
 
 source "googlecompute" "gpu-ubuntu" {
-  image_name          = "skypilot-gcp-gpu-ubuntu-${local.timestamp}"
+  image_name          = "skypilot-gcp-gpu-ubuntu-${local.date}"
   project_id          = "sky-dev-465"
   source_image_family = "ubuntu-2204-lts"
   zone                = var.zone

diff --git a/sky/provision/azure/instance.py b/sky/provision/azure/instance.py
@@ -311,30 +311,10 @@ def _create_vm(
         vm_name=vm_name,
         parameters=vm_instance,
     )
-    # poller.result() will block on async operation until it's done.
-    logger.info(f'Created VM {vm_poller.result().name}.')
-    # Configure driver extension for A10 GPUs. A10 GPUs requires a
-    # special type of drivers which is available at Microsoft HPC
-    # extension. Reference:
-    # https://forums.developer.nvidia.com/t/ubuntu-22-04-installation-driver-error-nvidia-a10/285195/2
-    # This can take more than 20mins for setting up the A10 GPUs
-    if node_config.get('need_nvidia_driver_extension', False):
-        ext_poller = compute_client.virtual_machine_extensions.\
-            begin_create_or_update(
-            resource_group_name=provider_config['resource_group'],
-            vm_name=vm_name,
-            vm_extension_name='NvidiaGpuDriverLinux',
-            extension_parameters=compute.VirtualMachineExtension(
-                location=provider_config['location'],
-                publisher='Microsoft.HpcCompute',
-                type_properties_type='NvidiaGpuDriverLinux',
-                type_handler_version='1.9',
-                auto_upgrade_minor_version=True,
-                settings='{}'))
-        logger.info(
-            f'Created VM extension {ext_poller.result().name} for VM {vm_name}.'
-        )
-    return vm_poller.result()
+    # This line will block until the VM is created or the operation times out.
+    vm = vm_poller.result()
+    logger.info(f'Created VM {vm.name}.')
+    return vm
 
 
 def _create_instances(compute_client: 'azure_compute.ComputeManagementClient',

diff --git a/sky/templates/azure-ray.yml.j2 b/sky/templates/azure-ray.yml.j2
@@ -83,7 +83,6 @@ available_node_types:
               {%- for cmd in cloud_init_setup_commands %}
               {{ cmd }}
               {%- endfor %}
-        need_nvidia_driver_extension: {{need_nvidia_driver_extension}}
         {%- if disk_performance_tier is not none %}
         disk_performance_tier: {{disk_performance_tier}}
         {%- endif %}