Google has published a detailed explanation of how this will be achieved.
Google's Android LLVM toolchain team has announced the introduction of a feature called Automatic Feedback-Directed Optimization (AutoFDO) into the Android kernel itself, which could bring a huge leap in system performance.
AutoFDO, first introduced in Android 12 (2021), is a sampling-based optimization technique.
It uses real-world execution data, collected directly from hardware monitors, to guide compiler decisions.
In this way, static analysis is replaced with profiles that reflect the actual usage of the device.
This feature is similar to the PGO (Profile Guided Optimization) technology already used in Windows and Linux, as well as in Google's Chromium-based applications.
According to Google, this approach allows the compiler to more easily identify "hot" code paths - those that are executed most often - and optimize them accordingly.
These profiles are created under controlled laboratory conditions by running representative workloads, including the 100 most popular applications.
In its simplest terms, the principle is somewhat similar to how data caching works.
Google states that their research shows an average performance increase of 10.5 percent.
Considering that the Android kernel takes up about 40% of the total processor (CPU) time, the results are tangible: 4% faster cold launch of applications and 1% faster boot time.
While these percentages don't seem huge on paper, Google claims that these optimizations are great for the overall user experience as they translate into faster switching between apps, a more fluid interface, and longer battery life.
The feature is currently targeting kernel versions android16-6.12 and android15-6.6, and data was collected on Pixel devices. Google plans to expand support to newer versions of Generic Kernel Images (GKI) and other architectures beyond the current aarch64.
By adopting a strategy that is “conservative by default,” Google points out that it is trying to ensure that AutoFDO increases speed without compromising the reliability of the system itself.
Comments