Support spot for GCP, and add spot info to Azure catalog (#113)

* Support spot instance for GCP and Azure * Fix Azure zones * Add Spot info in Azure catalog
skypilot-org · Dec 20, 2021 · f4c858d · f4c858d
1 parent 0192eb4
commit f4c858d
Show file tree

Hide file tree

Showing 8 changed files with 3,431 additions and 3,206 deletions.
diff --git a/prototype/config/azure-ray.yml.j2 b/prototype/config/azure-ray.yml.j2
@@ -30,6 +30,13 @@ available_node_types:
             imageOffer: ubuntu-1804
             imageSku: 1804-gen2
             imageVersion: 21.09.13
+            # optionally set priority to use Spot instances
+            {%- if use_spot %}
+            priority: Spot
+            # set a maximum price for spot instances if desired
+            # billingProfile:
+            #     maxPrice: -1
+            {%- endif %}
         # TODO: attach disk
 {% if num_nodes > 1 %}
     ray.worker.default:
@@ -44,6 +51,12 @@ available_node_types:
             imageOffer: ubuntu-1804
             imageSku: 1804-gen2
             imageVersion: 21.09.13
+          {%- if use_spot %}
+            priority: Spot
+            # set a maximum price for spot instances if desired
+            # billingProfile:
+            #     maxPrice: -1
+          {%- endif %}
 {%- endif %}
 
 head_node_type: ray.head.default

diff --git a/prototype/config/gcp-ray.yml.j2 b/prototype/config/gcp-ray.yml.j2
@@ -43,6 +43,9 @@ available_node_types:
           - key: install-nvidia-driver
             value: "True"
       scheduling:
+{%- if use_spot %}
+        - preemptible: true
+{%- endif %}
         - onHostMaintenance: TERMINATE  # Required for GPU-attached VMs.
 {%- endif %}
 {% if num_nodes > 1 %}
@@ -71,6 +74,9 @@ available_node_types:
           - key: install-nvidia-driver
             value: "True"
       scheduling:
+{%- if use_spot %}
+        - preemptible: true
+{%- endif %}
         - onHostMaintenance: TERMINATE  # Required for GPU-attached VMs.
 {%- endif %}
 {%- endif %}

diff --git a/prototype/sky/backends/cloud_vm_ray_backend.py b/prototype/sky/backends/cloud_vm_ray_backend.py
@@ -276,7 +276,8 @@ def _update_blocklist_on_azure_error(self, region, zones, stdout, stderr):
         errors = [
             s.strip()
             for s in stdout_splits + stderr_splits
-            if 'Exception Details:' in s.strip()
+            if ('Exception Details:' in s.strip() or
+                'InvalidTemplateDeployment' in s.strip())
         ]
         if not errors:
             logger.info('====== stdout ======')

diff --git a/prototype/sky/clouds/azure.py b/prototype/sky/clouds/azure.py
@@ -85,7 +85,7 @@ def get_accelerators_from_instance_type(
     def make_deploy_resources_variables(self, task):
         r = task.best_resources
         assert not r.use_spot, \
-            'We currently do not support spot instances for Azure.'
+            'Our subscription offer ID does not support spot instances.'
         # r.accelerators is cleared but .instance_type encodes the info.
         acc_dict = self.get_accelerators_from_instance_type(r.instance_type)
         if acc_dict is not None:
@@ -95,6 +95,7 @@ def make_deploy_resources_variables(self, task):
         return {
             'instance_type': r.instance_type,
             'custom_resources': custom_resources,
+            'use_spot': r.use_spot,
         }
 
     def get_feasible_launchable_resources(self, resources):

diff --git a/prototype/sky/clouds/gcp.py b/prototype/sky/clouds/gcp.py
@@ -72,6 +72,64 @@ class GCP(clouds.Cloud):
     _ON_DEMAND_PRICES.update(_ON_DEMAND_PRICES_GPUS)
     _ON_DEMAND_PRICES.update(_ON_DEMAND_PRICES_TPUS)
 
+    _SPOT_PRICES = {
+        # VMs: https://cloud.google.com/compute/all-pricing.
+        # N1 standard
+        'n1-standard-1': 0.01,
+        'n1-standard-2': 0.02,
+        'n1-standard-4': 0.04,
+        'n1-standard-8': 0.08,
+        'n1-standard-16': 0.16,
+        'n1-standard-32': 0.32,
+        'n1-standard-64': 0.64,
+        'n1-standard-96': 0.96,
+        # N1 highmem
+        'n1-highmem-2': 0.024906,
+        'n1-highmem-4': 0.049812,
+        'n1-highmem-8': 0.099624,
+        'n1-highmem-16': 0.199248,
+        'n1-highmem-32': 0.398496,
+        'n1-highmem-64': 0.796992,
+        'n1-highmem-96': 1.195488,
+    }
+    # GPUs: https://cloud.google.com/compute/gpus-pricing.
+    _SPOT_PRICES_GPUS = {
+        # T4
+        'T4': 0.11,
+        '1x T4': 0.11,
+        '2x T4': 0.11 * 2,
+        '4x T4': 0.11 * 4,
+        # P4
+        'P4': 0.216,
+        '1x P4': 0.216,
+        '2x P4': 0.216 * 2,
+        '4x P4': 0.216 * 4,
+        # V100
+        'V100': 0.74,
+        '1x V100': 0.74,
+        '2x V100': 0.74 * 2,
+        '4x V100': 0.74 * 4,
+        '8x V100': 0.74 * 8,
+        # P100
+        'P100': 0.43,
+        '1x P100': 0.43,
+        '2x P100': 0.43 * 2,
+        '4x P100': 0.43 * 4,
+        # K80
+        'K80': 0.0375,
+        '1x K80': 0.0375,
+        '2x K80': 0.0375 * 2,
+        '4x K80': 0.0375 * 4,
+        '8x K80': 0.0375 * 8,
+    }
+    # TPUs: https://cloud.google.com/tpu/pricing.
+    _SPOT_PRICES_TPUS = {
+        'tpu-v2-8': 1.35,
+        'tpu-v3-8': 2.40,
+    }
+    _SPOT_PRICES.update(_SPOT_PRICES_GPUS)
+    _SPOT_PRICES.update(_SPOT_PRICES_TPUS)
+
     #### Regions/Zones ####
 
     @classmethod
@@ -125,9 +183,8 @@ def region_zones_provision_loop(
     #### Normal methods ####
 
     def instance_type_to_hourly_cost(self, instance_type, use_spot):
-        # TODO: use_spot support
         if use_spot:
-            return clouds.Cloud.UNKNOWN_COST
+            return GCP._SPOT_PRICES[instance_type]
         return GCP._ON_DEMAND_PRICES[instance_type]
 
     def accelerators_to_hourly_cost(self, accelerators):
@@ -168,15 +225,14 @@ def get_default_region(cls) -> clouds.Region:
 
     def make_deploy_resources_variables(self, task):
         r = task.best_resources
-        assert not r.use_spot, \
-            'We currently do not support spot instances for GCP'
         # Find GPU spec, if any.
         resources_vars = {
             'instance_type': r.instance_type,
             'gpu': None,
             'gpu_count': None,
             'tpu': None,
             'custom_resources': None,
+            'use_spot': r.use_spot,
         }
         accelerators = r.get_accelerators()
         if accelerators is not None:

diff --git a/prototype/sky/clouds/service_catalog/azure_catalog.py b/prototype/sky/clouds/service_catalog/azure_catalog.py
@@ -16,8 +16,9 @@ def get_hourly_cost(instance_type: str,
                     region: str = _DEFAULT_REGION,
                     use_spot: bool = False) -> float:
     """Returns the cost, or the cheapest cost among all zones for spot."""
-    assert not use_spot, 'not implemented'
-    return common.get_hourly_cost_impl(_df, instance_type, region, False)
+    # Ref: https://azure.microsoft.com/en-us/support/legal/offer-details/
+    assert not use_spot, 'Current Azure subscription does not support spot.'
+    return common.get_hourly_cost_impl(_df, instance_type, region, use_spot)
 
 
 def get_accelerators_from_instance_type(instance_type: str