David Mytton

Measuring website energy consumption via browser profiling

Published (updated: ) in Environment, IT Energy. Tags: , .

Measuring website energy consumption via browser profiling

Network energy consumption is not proportional to data transferred. The theory behind this is demonstrated in a conference paper, Malmodin (2020), and applied to video streaming in Carbon Trust, 2021. We can see this from statistics published by telecoms operators showing their network energy consumption decreased even as data volumes grew.

kWh/GB energy intensity values represent the average energy per unit based on knowing the total energy consumption and the total data transferred. This method is useful for allocating energy across the entire network when the total energy and total data transfer are known, but produces misleading results if used to measure how energy consumption changes based on a change in data volume.

Therefore, website carbon emissions calculators are not useful because they use an invalid method of calculating energy consumption.

Annual energy use and network data flows for two large network providers, expressed as an index relative to 2016 = 1.0. Source: Koomey & Masanet, 2021.

The network is also not the largest source of energy consumption when looking at the system end to end. The user device is often a large part of the environmental footprint of a particular application. This varies significantly by device and also by application, but to take video streaming as a common application, the user device makes up ~50% of the total energy profile.

Breakdown of emissions and energy consumption by video streaming process component for Europe in 2020. Source: The Carbon Trust, 2021.

This means that understanding the energy consumption of the user device is important to being able to optimize their efficiency.

The browser as the power profiler

With so many applications accessed through the web browser, profiling applications has become an important part of developing software. Every browser has its own set of devtools which allow inspecting the source code, analyzing network requests, profiling memory & CPU usage, and various other things.

Safari has had a power meter available in the developer tools for a long time, but it is displayed as a gauge that shows the “energy impact” in terms of “low”, “medium”, and “high”. It primarily focuses on CPU and recommends that idle pages keep CPU at less than 3%. This is directionally useful, but doesn’t offer any numbers. We really need actual numbers to be able to make comparisons and track trends.

Screenshot of the Safari developer tools showing the energy impact gauge.

Profiling website power consumption in Firefox

Firefox 104 (Aug 24 2022) introduced a new component in the profiler which reports power usage in watts (thanks to the Green Coding blog for highlighting this new functionality). This is available by querying power APIs made available through Windows 11 or Windows 10 (SDK 10.0.17763 (Windows 10 1809 / Redstone 5) and later) and macOS on Apple Silicon devices.

On Windows, the Energy Metering Interface (EMI) allows drivers to expose energy consumption data. This returns the total energy consumption up to the present, so can be queried multiple times by a profiler to determine the energy consumption during the period. The Firefox code can be seen doing this:

void AddSample(ULONGLONG aAbsoluteEnergy, ULONGLONG aAbsoluteTime) {
    // aAbsoluteTime is the time since the system start in 100ns increments.
    if (aAbsoluteTime == mPreviousTime) {
      return;
    }

    if (aAbsoluteEnergy > mPreviousValue) {
      int64_t increment = aAbsoluteEnergy - mPreviousValue;
      mCounter += increment;
      mPreviousValue += increment;
      mPreviousTime = aAbsoluteTime;
    }
}

On macOS, the OS provides an API to return the energy consumption of a specific task in nano-joules. Firefox converts this to picowatt-hours (this conversion is important because joules represent the energy whereas picowatt-hours represents the power consumption. “Power is the rate at which something uses energy.” Source: Sustainable energy without the hot air) for consistency with Windows:

private:
  int64_t GetTaskEnergy() {
    task_power_info_v2_data_t task_power_info;
    mach_msg_type_number_t count = TASK_POWER_INFO_V2_COUNT;
    kern_return_t kr = task_info(mach_task_self(), TASK_POWER_INFO_V2,
                                 (task_info_t)&task_power_info, &count);
    if (kr != KERN_SUCCESS) {
      return 0;
    }

    // task_energy is in nanojoules. To be consistent with the Windows EMI
    // API, return values in picowatt-hour.
    return task_power_info.task_energy / 3.6;
  }

The result is we can profile the power consumption of a website. You can separate the various processes which means you can isolate the website or application from Firefox itself and the various web rendering processes. This includes the decoder component when streaming videos, which is important to be able to understand what is consuming the most energy.

Isolating each component allows the developer to focus on the code they control without the profiling metrics being impacted by things out-of-scope, such as background processes and the OS itself.

Update 2022-09-06: A change was merged into the Firefox Profiler UI which now displays the sum of the power consumption in the selected time range.

Screenshot of a power profile capture in Firefox 104 on macOS.

What about other browsers?

Unfortunately, neither Safari nor Firefox are popular browsers on desktop (Statcounter shows desktop share at 66% Chrome, 10% Edge, 9% Safari, and 8% Firefox as of Sept 2022) so the question is if/when will this appear in Chrome (and then Brave, etc)?

There is currently an open proposal for “sustainability” metrics in Microsoft Edge, but this is different from showing power consumption. My feedback is that energy metrics are more useful than a carbon value or a sustainability score because they are more suitable for optimization by developers. Let’s see what happens.