- AS a developer, I want to estimate the carbon impact of my application (which uses several infrastructure components)
- As a developer / admin, I want to be able to see the carbon impact of my application over time (in dashboards)
This project is a POC for an Azure SCI Metrics framework (as in Software carbon intensity) based on the spec from the https://github.com/Green-Software-Foundation/ief/
This project is an implementation of the Impact engine Framework, that provides :
- An API : to calculate carbon metrics of an application (ad-hoc)
- A metrics exporter : which allows to build impact dashboards to help reduce carbon impact
- A python implementation of the framework
- An SCI model for compute server, based on the SCI study case : https://github.com/Green-Software-Foundation/sci-guide/blob/dev/use-case-submissions/dow-msft-Graph-DB.md
- Impotant : the current implementation is a work in progress to converge with latest spec from the GSF IEF project.
The impactModel is the low level object that defines the impact calculation (here SCI), and the impact metrics that are used in the equation (CPU, RAM, GPU + timespan for SCI)
The impactNode is the object that represents the infrastructure component (VM, pod, etc..) ; it is associated to an impactModel. It fetches the metrics from the infrastructure component, and calculates the model.calculate to return an imapctMetric.
the user creates an impactRequest (Yaml, json) that describes the application, and the resources it uses (VMs, pods, etc..) ; and the timespan over which he wants to get the carbon impact of the application (R in the SCI equation)
the user sends the impactRequest to the Impact Engine API or CLI, which returns the carbon impact of the application (SCI) ; and the carbon impact of each resource used by the application (E, I, M)
-
Model : SCI based on the DOW study case (can be used for any compute server)
-
Components (or impactNodes) : AzureVM, AKSNode, AKSPod, AzureFunction -> all use the same model
In this example, our app e-shop, has a dedicated VM, and we want to get the carbon impact of the application over the last 24 hour (timespan).
so SCI of e-shop App = SCI of VM (because the VM is not shared with other apps)
curl -X POST \
http://localhost:8000/metrics \
-H 'Content-Type: application/json' \
-d '{
"app_name": "e-shop",
"components": [
{
"name": "webserver",
"auth_params": {
},
"type": "AzureVM",
"resource_selectors": {
"subscription_id": "<SUbID>",
"resource_group": "<RGName>",
"name": "tototatar"
},
"metadata": {
"your_metadata_key": "your_metadata_value"
}
}
],
"interval": "PT1M", # to normlize or homogenize the metrics collected from different sources ; this is a technical parameter ; TODO : remove from query
"timespan": "PT24H" # the R of SCI equation (time span)
}'{
"e-shop": {
"name": "e-shop",
"unit": "severalUnits",
"type": "aggregatedimpactnode",
"model": "sumofcomponents",
"description": "Description of SCI Impact Metrics",
"timespan": "PT24H",
"interval": "PT1M",
"E_CPU": 1.152,
"E_MEM": 0.6218425726807492,
"E_GPU": 0.0,
"E": 1.7738425726807492,
"I": 100.0,
"M": 4.280821917808219e-05,
"SCI": 177.3843000762941,
"metadata": {
"aggregated": "True"
},
"observations": {},
"components": [{
"tototatar": {
"name": "tototatar",
"unit": "severalUnits",
"type": "azurevm",
"model": "computeserver_static_imp",
"description": "Description of SCI Impact Metrics",
"timespan": "PT24H",
"interval": "PT1M",
"E_CPU": 1.152,
"E_MEM": 0.6218425726807492,
"E_GPU": 0.0,
"E": 1.7738425726807492,
"I": 100.0,
"M": 4.280821917808219e-05,
"SCI": 177.3843000762941,
"metadata": {
"resource_name": "tototatar"
},
"observations": {
"average_cpu_percentage": 2.4472038327526153,
"average_memory_gb": 1.6364278228440767,
"average_gpu_percentage": 0
},
"components": [],
"host_node": {}
}
}],
"host_node": {}
}
}In this example, our app e-shop, has a shared VM, and we want to get the carbon impact of the application over the last 24 hour (timespan).
auth_params = {
}
resource_selectors = {
"subscription_id": "<SUB>",
"resource_group": "<RG>",
"name": "<VMName>",
}
metadata = {
"region": "westeurope"
}
timespan = "PT24H"
interval = "PT5M"
vm = AzureVM(name = "mywebserver", model = ComputeServer_STATIC_IMP(),
carbon_intensity_provider=None,
auth_object=auth_params,
resource_selectors=resource_selectors,
metadata=metadata,
timespan=timespan,
interval=interval)
#print(vm.fetch_resources())
#print(vm.fetch_observations())
print(vm.calculate())
manual_observations = {
"node_host_cpu_util_percent" : 50,
"node_host_memory_util_percent" : 50,
"node_host_gpu_util_percent" : 50
}
workload = AttributedImpactNodeInterface(name = "myworkload",
host_node=vm,
carbon_intensity_provider=None,
metadata=metadata,
observations=manual_observations,
timespan=timespan,
interval=interval)
print(workload.calculate()){
'myworkload': SCIImpactMetricsInterface(name = 'myworkload', unit = 'severalUnits', type = 'attributedimpactnode', model = 'attributedimpactfromnode', description = 'Description of SCI Impact Metrics', timespan = 'PT24H', interval = 'PT5M', E_CPU = 0.576, E_MEM = 0.31337428009815704, E_GPU = 0.0, E = 0.889374280098157, I = 100.0, M = 4.280821917808219e-05, SCI = 88.93747081803488, metadata = {
'attributed': 'True',
'host_node_name': 'mywebserver'
}, observations = {
'node_host_cpu_util_percent': 50,
'node_host_memory_util_percent': 50,
'node_host_gpu_util_percent': 50
}, components = [], host_node = {
'mywebserver': SCIImpactMetricsInterface(name = 'tototatar', unit = 'severalUnits', type = 'azurevm', model = 'computeserver_static_imp', description = 'Description of SCI Impact Metrics', timespan = 'PT24H', interval = 'PT5M', E_CPU = 1.152, E_MEM = 0.6267485601963141, E_GPU = 0.0, E = 1.778748560196314, I = 100.0, M = 4.280821917808219e-05, SCI = 177.8748988278506, metadata = {
'resource_name': 'tototatar'
}, observations = {
'average_cpu_percentage': 2.9743745726495727,
'average_memory_gb': 1.6493383163060897,
'average_gpu_percentage': 0
}, components = [], host_node = {})
})
}In this example, our App is deployed on an AKS cluster, and we want to get the carbon impact of the application over the last 24 hour (timespan).
we define resource selectors to select the resources we want to get the carbon impact of. In this case, we want to get the carbon impact of the keda namespace, which contains the keda operator and the keda metrics server.
we also define the prometheus endpoint, which is the endpoint of the prometheus server that is used by the keda metrics server to get metrics from the cluster.
Our App is composed in this case of 2 pods, each pod running on a node (could be same node, or different one)
The Imapct of a pod, is deducated from the impact of the node it is running on. (attribution pattern) ; and we use % pod utl of node resouces ( average % CPU, % of memory and % of GPU) -> during the timespan defined by the query
so :
- SCI of e-shop App = sum of SCI of pods ;
- and SCI of pod = SCI of node * % of node resources used by the pod
curl -X POST \
http://localhost:8000/metrics \
-H 'Content-Type: application/json' \
-d '{
"app_name": "e-shop",
"components": [
{
"name": "batchprocessing",
"auth_params": {
"key": "value"
},
"type": "AKSPod",
"resource_selectors": {
"subscription_id": "<SUB>",
"resource_group": "<RG>",
"cluster_name": "<cluster>",
"namespace": "keda",
"prometheus_endpoint": "<Prom endpoint>"
},
"metadata": {
"your_metadata_key": "your_metadata_value"
}
}
],
"interval": "PT5M", # to normlize or homogenize the metrics collected from different sources ; this is a technical parameter ;
"timespan": "PT24H" # the R of SCI equation (time span)
}'{
"e-shop": {
"name": "e-shop",
"unit": "severalUnits",
"type": "aggregatedimpactnode",
"model": "sumofcomponents",
"description": "Description of SCI Impact Metrics",
"timespan": "PT24H",
"interval": "PT5M",
"E_CPU": 1.0295646315789783e-05,
"E_MEM": 0.0,
"E_GPU": 0.0,
"E": 1.0295646315789783e-05,
"I": 100.0,
"M": 3.567351598173516e-06,
"SCI": 0.0010331319831771517,
"metadata": {
"aggregated": "True"
},
"observations": {},
"components": [{
"keda-metrics-apiserver-5bcc58d658-mxqh7": {
"name": "keda-metrics-apiserver-5bcc58d658-mxqh7",
"unit": "severalUnits",
"type": "attributedimpactnode",
"model": "attributedimpactfromnode",
"description": "Description of SCI Impact Metrics",
"timespan": "PT1H",
"interval": "PT5M",
"E_CPU": 7.680000000002207e-06,
"E_MEM": 0.0,
"E_GPU": 0.0,
"E": 7.680000000002207e-06,
"I": 100.0,
"M": 1.783675799086758e-06,
"SCI": 0.0007697836757993074,
"metadata": {
"attributed": "True",
"host_node_name": "aks-agentpool-23035252-vmss000005"
},
"observations": {
"node_host_cpu_util_percent": 0.006000000000001724,
"node_host_memory_util_percent": 0.046112060546875,
"node_host_gpu_util_percent": 0
},
"components": [],
"host_node": {
"aks-agentpool-23035252-vmss000005": {
"name": "aks-agentpool-23035252-vmss000005",
"unit": "severalUnits",
"type": "azurevm",
"model": "computeserver_static_imp",
"description": "Description of SCI Impact Metrics",
"timespan": "PT1H",
"interval": "PT5M",
"E_CPU": 0.128,
"E_MEM": 0.0,
"E_GPU": 0.0,
"E": 0.128,
"I": 100.0,
"M": 1.783675799086758e-06,
"SCI": 12.8000017836758,
"metadata": {
"resource_name": "aks-agentpool-23035252-vmss000005"
},
"observations": {
"average_cpu_percentage": 12.436124999999999,
"average_memory_gb": 0,
"average_gpu_percentage": 0
},
"components": [],
"host_node": {}
}
}
},
"keda-operator-778fb74497-m55gs": {
"name": "keda-operator-778fb74497-m55gs",
"unit": "severalUnits",
"type": "attributedimpactnode",
"model": "attributedimpactfromnode",
"description": "Description of SCI Impact Metrics",
"timespan": "PT1H",
"interval": "PT5M",
"E_CPU": 2.615646315787576e-06,
"E_MEM": 0.0,
"E_GPU": 0.0,
"E": 2.615646315787576e-06,
"I": 100.0,
"M": 1.783675799086758e-06,
"SCI": 0.00026334830737784437,
"metadata": {
"attributed": "True",
"host_node_name": "aks-agentpool-23035252-vmss000005"
},
"observations": {
"node_host_cpu_util_percent": 0.0020434736842090437,
"node_host_memory_util_percent": 0.038562774658203125,
"node_host_gpu_util_percent": 0
},
"components": [],
"host_node": {
"aks-agentpool-23035252-vmss000005": {
"name": "aks-agentpool-23035252-vmss000005",
"unit": "severalUnits",
"type": "azurevm",
"model": "computeserver_static_imp",
"description": "Description of SCI Impact Metrics",
"timespan": "PT1H",
"interval": "PT5M",
"E_CPU": 0.128,
"E_MEM": 0.0,
"E_GPU": 0.0,
"E": 0.128,
"I": 100.0,
"M": 1.783675799086758e-06,
"SCI": 12.8000017836758,
"metadata": {
"resource_name": "aks-agentpool-23035252-vmss000005"
},
"observations": {
"average_cpu_percentage": 12.436124999999999,
"average_memory_gb": 0,
"average_gpu_percentage": 0
},
"components": [],
"host_node": {}
}
}
}
}],
"host_node": {}
}
}