Teads CPU pipeline

The Teads CPU power curve CPU utilization (as a percentage) against a scaling factor that can be applied to the CPUs thermal design power to estimate the power drawn by the CPU in Watts.

The research underpinning the curve was summarized in a pair of blog posts:

TEADS Engineering: Buildiong an AWS EC2 Carbon Emissions Dataset

Teads Engineering: Estimating AWS EC2 Instances Power Consumption

The curve has become very widely used as a general purpose utilization-to-wattage converter for CPUs, despite the fact that it does not geenralize well.

The wattage can be transformed into energy by doing the following:

1. Measure your CPU utilization
2. Determine the thermal design power of your processor
3. Determine the scaling factor for your CPU utilization by interpolating the Teads curve
4. Determine the power drawn by your CPU by multiplying your scaling factor by the CPU's thermal design power
5. Perform a unit conversion to convert power in Watts to energy in kwH
6. Scale the energy estimated for the entire chip to the portion of the chip that is actually in use.

These steps can be executed in IF using just three plugins:

• Interpolate
• Multiply
• Divide

We'll go through each step in the energy estimate and examine how to implement it in a manifest file using IF's standard library of builtins.

Impact Framework implementation

First, create a manifest file and add this following boilerplate code:

name: carbon-intensity plugin demo
description:
tags:
initialize:
  plugins:
tree:
  children:
    child:
      pipeline:
        observe:
        regroup:
        compute:
      defaults:
      inputs:

If this structure looks unfamiliar to you, you can go back to our manifests page.

Step 1: measure CPU utilization

The first step was to measure your CPU utilization. In real use cases you would typically do this using an importer plugin that grabs data from a monitor API or similar. However, for this example we will just manually create some dummy data. Add some timestamps, durations and cpu/utilization data to your inputs array, as follows:

name: teads demo
description:
tags:
initialize:
  plugins:
tree:
  children:
    child:
      pipeline:
        observe:
        regroup:
        compute:
      defaults:
      inputs:
        - timestamp: 2023-08-06T00:00
          duration: 360
          cpu/utilization: 1
          carbon: 30
        - timestamp: 2023-09-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 10
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 50
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 100

Step 2: Determine the thermal design power of your processor

Typically determinign the TDP of your processor would be done using a CSV lookup. For now, we will just hard code some TDP data into your manifest so we can focus on the CPU utilization to energy calculations. Add thermal-design-power to defaults - this is a shortcut to providing it in every timestep in your inputs array.

default:
  thermal-design-power: 100

Step 3: Interpolate the Teads curve

The Teads curve has CPU utilization on the x axis and a scaling factor on the y axis. There are only four points on the published curve. Your task is to get the scaling factor for your specific CPU utilization values by interpolating between the known points. Luckily, we have a builtin for that purpose!

Add the Interpolation plugin to your list of plugins in the initialize block.

initialize:
  plugins:
    interpolate:
      method: Interpolation
      path: builtin

The details about the interpolation you want to do and the values to return are configured in the config which is also added in the initialize block. Specifically, you have to provide the known points of the curve you want to interpolate, the input-parameter (which is the x value whose correspondiong y value you want to find out, i.e. your CPU utilization value) and the output-parameter (the name you want to give to your retrieved y value).

You want to interpolate the Teads curve, so you can provide the x and y values obtained from the articles linked in the introduction section above:

x: [0, 10, 50, 100]
y: [0.12, 0.32, 0.75, 1.02]

Your input-parameter is your cpu/utilization and we'll name give the output-parameter the name cpu-factor.

Your compelted initialize block for interpolate should look as follows:

interpolate:
  method: Interpolation
  path: 'builtin'
  config:
    method: linear
    x: [0, 10, 50, 100]
    y: [0.12, 0.32, 0.75, 1.02]
    input-parameter: 'cpu/utilization'
    output-parameter: 'cpu-factor'

Step 4: Convert CPU factor to power

The interpoaltion only gives us the scaling factor; we need to apply that scaling factor to the processor's TDP to get the power drawn by the CPU at your specific CPU utilization.

To do this, we can use the Multiply plugin in the IF standard library. We'll give the instance of Multiply the name cpu-factor-to-wattage and in the config we'll define cpu-factor and thermal-design-power as the two elements in our inputs array that we want to multiply together. Then we'll name the result cpu-wattage:

cpu-factor-to-wattage:
  method: Multiply
  path: builtin
  config:
    input-parameters: ['cpu-factor', 'thermal-design-power']
    output-parameter: 'cpu-wattage'

Add this to your initialize block.

Step 5: Convert wattage to energy

Next we have to perform some unit conversions. Wattage is a measure of power (energy over time). To convert to energy, we can first multiply by the number of seconds our observation covers (duration) to yield energy in joules. Then, convert to kWh by applying a scaling factor that takes seconds to hours and watts to kilowatts.

You can do this in two steps: the first uses another instance of Multiply an the second uses Divide:

To do the initial multiplication of the CPU wattage and the observation duration, add the following config to your initialize block:

wattage-times-duration:
  method: Multiply
  path: builtin
  config:
    input-parameters: ['cpu-wattage', 'duration']
    output-parameter: 'cpu-wattage-times-duration'

next, use the Divide plugin to do the unit conversion:

wattage-to-energy-kwh:
  method: Divide
  path: 'builtin'
  config:
    numerator: cpu-wattage-times-duration
    denominator: 3600000
    output: cpu-energy-raw

Step 6: Scale the energy by the allocated CPUs

The cpu-energy-raw value you just configured is for the entire chip. But your application probably doesn't use the entire chip. Chances are you have some number of VCPUs allocated to you that is less than the total available. So you can scale your energy estimate by the ratio of VCPUs allocated to VCPUS available.

Let's assume you know the number of VCPUs allocated and available in advance and that they are the same in every timestep. In this case, you can just add the values to defaults so they become available in every timestep, just as you did with thermal-design-power.

defaults:
  thermal-design-power: 100
  vcpus-total: 8
  vcpus-allocated: 2

You need one instance of Divide to calculate the vcpu-ratio and another to apply that vcpu-ratio to your cpu-energy-raw value and yield your final result: cpu-energy-kwh. Add the following to your initialize block to achieve those steps:

calculate-vcpu-ratio:
  method: Divide
  path: 'builtin'
  config:
    numerator: vcpus-total
    denominator: vcpus-allocated
    output: vcpu-ratio
correct-cpu-energy-for-vcpu-ratio:
  method: Divide
  path: 'builtin'
  config:
    numerator: cpu-energy-raw
    denominator: vcpu-ratio
    output: cpu-energy-kwh

Step 7: Define your pipeline

Now you have configured all your plugins, covering all the stages of the calculation, you can simple define them in order in the pipeline section of your manifest, as follows:

tree:
  children:
    child:
      pipeline:
        observe:
        regroup:
        compute:
          - interpolate
          - cpu-factor-to-wattage
          - wattage-times-duration
          - wattage-to-energy-kwh
          - calculate-vcpu-ratio
          - correct-cpu-energy-for-vcpu-ratio

You also need to add some input data that your pipeline can operate over.

You can see the full manifest in the IF repository.

That's it! Your manifest is ready to run!

Running the manifest

Having saved your manifest as teads-curve.yaml you can run it using IF:

if-run -m teads-curve.yml -o teads-output.yml

This will yield the following output file:

name: teads curve demo
description: null
tags: null
initialize:
  plugins:
    interpolate:
      path: builtin
      method: Interpolation
      config:
        method: linear
        x:
          - 0
          - 10
          - 50
          - 100
        'y':
          - 0.12
          - 0.32
          - 0.75
          - 1.02
        input-parameter: cpu/utilization
        output-parameter: cpu-factor
    cpu-factor-to-wattage:
      path: builtin
      method: Multiply
      config:
        input-parameters:
          - cpu-factor
          - thermal-design-power
        output-parameter: cpu-wattage
    wattage-times-duration:
      path: builtin
      method: Multiply
      config:
        input-parameters:
          - cpu-wattage
          - duration
        output-parameter: cpu-wattage-times-duration
    wattage-to-energy-kwh:
      path: builtin
      method: Divide
      config:
        numerator: cpu-wattage-times-duration
        denominator: 3600000
        output: cpu-energy-raw
    calculate-vcpu-ratio:
      path: builtin
      method: Divide
      config:
        numerator: vcpus-total
        denominator: vcpus-allocated
        output: vcpu-ratio
    correct-cpu-energy-for-vcpu-ratio:
      path: builtin
      method: Divide
      config:
        numerator: cpu-energy-raw
        denominator: vcpu-ratio
        output: cpu-energy-kwh
execution:
  command: >-
    /home/user/.npm/_npx/1bf7c3c15bf47d04/node_modules/.bin/ts-node
    /home/user/if/src/index.ts -m manifests/examples/teads-curve.yml
  environment:
    if-version: 0.6.0
    os: macOS
    os-version: 14.6.1
    node-version: 18.20.4
    date-time: 2024-10-03T15:05:11.948Z (UTC)
    dependencies:
      - '@babel/core@7.22.10'
      - '@babel/preset-typescript@7.23.3'
      - '@commitlint/cli@18.6.0'
      - '@commitlint/config-conventional@18.6.0'
      - '@grnsft/if-core@0.0.25'
      - '@jest/globals@29.7.0'
      - '@types/jest@29.5.8'
      - '@types/js-yaml@4.0.9'
      - '@types/luxon@3.4.2'
      - '@types/node@20.9.0'
      - axios-mock-adapter@1.22.0
      - axios@1.7.2
      - cross-env@7.0.3
      - csv-parse@5.5.6
      - csv-stringify@6.4.6
      - fixpack@4.0.0
      - gts@5.2.0
      - husky@8.0.3
      - jest@29.7.0
      - js-yaml@4.1.0
      - lint-staged@15.2.2
      - luxon@3.4.4
      - release-it@16.3.0
      - rimraf@5.0.5
      - ts-command-line-args@2.5.1
      - ts-jest@29.1.1
      - typescript-cubic-spline@1.0.1
      - typescript@5.2.2
      - winston@3.11.0
      - zod@3.23.8
  status: success
tree:
  children:
    child:
      pipeline:
        observe:
        regroup:
        compute:
          - interpolate
          - cpu-factor-to-wattage
          - wattage-times-duration
          - wattage-to-energy-kwh
          - calculate-vcpu-ratio
          - correct-cpu-energy-for-vcpu-ratio
      defaults:
        thermal-design-power: 100
        vcpus-total: 8
        vcpus-allocated: 2
      inputs:
        - timestamp: 2023-08-06T00:00
          duration: 360
          cpu/utilization: 1
          carbon: 30
        - timestamp: 2023-09-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 10
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 50
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 100
      outputs:
        - timestamp: 2023-08-06T00:00
          duration: 360
          cpu/utilization: 1
          carbon: 30
          thermal-design-power: 100
          vcpus-total: 8
          vcpus-allocated: 2
          cpu-factor: 0.13999999999999999
          cpu-wattage: 13.999999999999998
          cpu-wattage-times-duration: 5039.999999999999
          cpu-energy-raw: 0.0013999999999999998
          vcpu-ratio: 4
          cpu-energy-kwh: 0.00034999999999999994
        - timestamp: 2023-09-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 10
          thermal-design-power: 100
          vcpus-total: 8
          vcpus-allocated: 2
          cpu-factor: 0.32
          cpu-wattage: 32
          cpu-wattage-times-duration: 11520
          cpu-energy-raw: 0.0032
          vcpu-ratio: 4
          cpu-energy-kwh: 0.0008
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 50
          thermal-design-power: 100
          vcpus-total: 8
          vcpus-allocated: 2
          cpu-factor: 0.75
          cpu-wattage: 75
          cpu-wattage-times-duration: 27000
          cpu-energy-raw: 0.0075
          vcpu-ratio: 4
          cpu-energy-kwh: 0.001875
        - timestamp: 2023-10-06T00:00
          duration: 360
          carbon: 30
          cpu/utilization: 100
          thermal-design-power: 100
          vcpus-total: 8
          vcpus-allocated: 2
          cpu-factor: 1.02
          cpu-wattage: 102
          cpu-wattage-times-duration: 36720
          cpu-energy-raw: 0.0102
          vcpu-ratio: 4
          cpu-energy-kwh: 0.00255