Posted by Yabin Cui, Software program Engineer
We’re the Android LLVM toolchain workforce. Considered one of our prime priorities is to enhance Android efficiency by optimization strategies within the LLVM ecosystem. We’re continually trying to find methods to make Android sooner, smoother, and extra environment friendly. Whereas a lot of our optimization work occurs in userspace, the kernel stays the guts of the system. Immediately, we’re excited to share how we’re bringing Computerized Suggestions-Directed Optimization (AutoFDO) to the Android kernel to ship important efficiency wins for customers.
Throughout a normal software program construct, the compiler makes 1000’s of small selections, equivalent to whether or not to inline a perform and which department of a conditional is more likely to be taken, based mostly on static code hints.Whereas these heuristics are helpful, they do not at all times precisely predict code execution throughout real-world cellphone utilization.
AutoFDO modifications this through the use of real-world execution patterns to information the compiler. These patterns characterize the commonest instruction execution paths the code takes throughout precise use, captured by recording the CPU’s branching historical past. Whereas this knowledge may be collected from fleet units, for the kernel we synthesize it in a lab setting utilizing consultant workloads, equivalent to operating the highest 100 hottest apps. We use a sampling profiler to seize this knowledge, figuring out which components of the code are ‘sizzling’ (steadily used) and that are ‘chilly’.
Once we rebuild the kernel with these profiles, the compiler could make a lot smarter optimization selections tailor-made to precise Android workloads.
To grasp the impression of this optimization, contemplate these key information:
- On Android, the kernel accounts for about 40% of CPU time.
- We’re already utilizing AutoFDO to optimize native executables and libraries within the userspace, attaining about 4% chilly app launch enchancment and a 1% boot time discount.
Actual-World Efficiency Wins
Now we have seen spectacular enhancements throughout key Android metrics by leveraging profiles from managed lab environments. These profiles have been collected utilizing app crawling and launching, and measured on Pixel units throughout the 6.1, 6.6, and 6.12 kernels.
Probably the most noticeable enhancements are listed beneath. Particulars on the AutoFDO profiles for these kernel variations may be discovered within the respective Android kernel repositories for android16-6.12 and android15-6.6 kernels.
These aren’t simply theoretical numbers. They translate to a snappier interface, sooner app switching, prolonged battery life, and an total extra responsive system for the tip person.
Our deployment technique entails a classy pipeline to make sure profiles keep related and efficiency stays steady.
Step 1: Profile Assortment
Whereas we depend on our inside take a look at fleet to profile userspace binaries, we shifted to a managed lab setting for the Generic Kernel Picture (GKI). Decoupling profiling from the system launch cycle permits for versatile, quick updates unbiased of deployed kernel variations. Crucially, checks verify that this lab-based knowledge delivers efficiency positive aspects similar to these from real-world fleets.
System-Large Monitoring: Capturing not solely foreground app actions, but additionally essential background workloads and inter-process communications
Validation: This synthesized workload exhibits an 85% similarity to execution patterns collected from our inside fleet.
Focused Information: By repeating these checks sufficiently, we seize high-fidelity execution patterns that precisely characterize real-world person interplay with the most well-liked purposes. Moreover, this extensible framework permits us to seamlessly combine further workloads and benchmarks to broaden our protection.
Step 2: Profile Processing
We post-process the uncooked hint knowledge to make sure it’s clear, efficient, and prepared for the compiler.
Aggregation: We consolidate knowledge from a number of take a look at runs and units right into a single system view.
- Conversion: We convert uncooked traces into the AutoFDO profile format, filtering out undesirable symbols as wanted.
Profile Trimming: We trim profiles to take away knowledge for “chilly” features, permitting them to make use of normal optimization. This prevents regressions in not often used code and avoids pointless will increase in binary dimension.
Step 3: Profile Testing
Earlier than deployment, profiles endure rigorous verification to make sure they ship constant efficiency wins with out stability dangers.
Profile & Binary Evaluation: We strictly examine the brand new profile’s content material (together with sizzling features, pattern counts, and profile dimension) towards earlier variations. We additionally use the profile to construct a brand new kernel picture, analyzing binaries to make sure that modifications to the textual content part are in step with expectations.
Efficiency Verification: We run focused benchmarks on the brand new kernel picture. This confirms that it maintains the efficiency enhancements established by earlier baselines.
Steady Updates
Code naturally “drifts” over time, so a static profile would finally lose its effectiveness. To take care of peak efficiency, we run the pipeline constantly to drive common updates:
- Common Refresh: We refresh profiles in Android kernel LTS branches forward of every GKI launch, guaranteeing each construct contains the newest profile knowledge.
Future Growth: We’re presently delivering these updates to the android16-6.12 and android15-6.6 branches and can increase help to newer GKI variations, such because the upcoming android17-6.18.
A standard query with profile-guided optimization is whether or not it introduces stability dangers. As a result of AutoFDO primarily influences compiler heuristics, equivalent to perform inlining and code structure, moderately than altering the supply code’s logic, it preserves the useful integrity of the kernel. This know-how has already been confirmed at scale, serving as a normal optimization for Android platform libraries, ChromeOS, and Google’s personal server infrastructure for years.
To additional assure constant conduct, we apply a “conservative by default” technique. Capabilities not captured in our high-fidelity profiles are optimized utilizing normal compiler strategies. This ensures that the “chilly” or not often executed components of the kernel behave precisely as they might in a normal construct, stopping efficiency regressions or sudden behaviors in nook circumstances.
We’re presently deploying AutoFDO throughout the android16-6.12 and android15-6.6 branches. Past this preliminary rollout, we see a number of promising avenues to additional improve the know-how:
Expanded Attain: We look ahead to deploying AutoFDO profiles to newer GKI kernel variations and extra construct targets past the present aarch64 help.
GKI Module Optimization: At the moment, our optimization is concentrated on the primary kernel binary (vmlinux). Increasing AutoFDO to GKI modules may deliver efficiency advantages to a bigger portion of the kernel subsystem.
Vendor Module Help: We’re additionally excited about supporting AutoFDO for vendor modules constructed utilizing the Driver Growth Equipment (DDK). With help already out there in our construct system (Kleaf) and profiling instruments (simpleperf), this permits distributors to use these identical optimization strategies to their particular {hardware} drivers.
Broader Profile Protection: There may be potential to gather profiles from a wider vary of Vital Person Journeys (CUJs) to optimize them.
By bringing AutoFDO to the Android kernel, we’re guaranteeing that the very basis of the OS is optimized for the best way you utilize your system day-after-day.