Android App Performance Optimization with Baseline Profiles

2 posts · Sep 2012 – Oct 2025

What this trail covers

Meta's Android apps have significantly improved performance through the extensive use of Android's Baseline Profiles feature. This has addressed challenges related to slow startups and post-startup user journeys by enabling profile-guided compiler and runtime optimizations. The company has developed infrastructure to generate and tune these profiles using user data, leading to performance improvements of up to 40% across critical metrics. This builds upon previous efforts like Interdex Ordering and the development of a bytecode instrumentation system for performance telemetry.

All Meta trails

Oct 2025Accelerating our Android apps with Baseline ProfilesThis post details Meta's extensive use of Android's Baseline Profiles feature to improve the performance of its Android applications. It explains the challenges of ART's default runtime optimizations (interpretation and JIT compilation) and how Meta's apps, with tens of thousands of classes loaded on startup, face significant performance degradation. The post contrasts Cloud Profiles with Baseline Profiles, highlighting the developer control offered by the latter. Meta has developed infrastructure to generate and tune Baseline Profiles using user data, leading to up to 40% performance improvements in critical metrics. This builds on prior work like Interdex Ordering and Redex, and now leverages ART's install-time optimizations for AOT compilation and app image creation.
Mar 2018Profilo: Understanding app performance in the wildThis post introduces Profilo, a new high-throughput, mobile-first performance tracing library for Android, and announces its open-source release. Profilo enables the collection of detailed telemetry at rates of 3000 events per second or more with minimal disruption. It features a powerful remote configuration system for fine-tuning telemetry collection and provides precise insights into performance regressions as small as tens of milliseconds of CPU time. Unique features include a custom Java stack unwinder that bypasses official Java APIs for more efficient and accurate stack trace collection, and the ability to capture 'atrace' telemetry (systrace) in production to analyze VM and framework behavior like garbage collection and I/O operations. The post also highlights how Profilo has improved root-cause analysis for issues like cold app startup and resource usage.
Oct 2017Mobile @Scale 2017 recapThis post details the use of Litho and Infer for multi-threaded rendering of the Facebook Android News Feed codebase, achieving scalable, safe, and performant migration. It highlights encountered issues, performance results, and provides guidance for Android engineers on using these libraries.
Mar 2016Dive into React Native performanceThis post details specific optimizations made to improve React Native app startup time, including "doing less" (cleanup require/Babel helpers, avoiding string copying/decoding, stripping DEV-only modules, server-generated event descriptions) and "scheduling" improvements (lazy requires, Relay incremental cache read, de-batching bridge calls, early UI flushing, lazy native modules loading, lazy touch bindings, deferring popular events query). The primary focus is on reducing the time from tapping an event button to events being visible on screen, achieving a twofold improvement on the iPhone 5.
Nov 2015Optimizing Facebook for iOS start timeThis post details optimizations for Facebook's iOS app start time, focusing on reducing cold start duration. Key contributions include: defining stable metrics for app start (tap icon to feed loaded), breaking down cold start into request time, network time, and response processing time, identifying initialization as a bottleneck, refactoring dependencies to allow concurrent feed request processing, implementing UDP priming to pre-fetch data and reduce TCP connection setup time, and optimizing story-fetching strategy from batch to a streaming-like 1+1+X approach. Additionally, non-critical app services were enqueued to run after app start completion to reduce contention during initial loading.
Advertisement
Oct 2015Performance instrumentation for Android appsThis post details the development of a bytecode instrumentation system for Android apps to collect performance telemetry. It addresses limitations of manual instrumentation and Android's built-in profilers by using a rule-based bytecode rewriter (ASM) to inject code at entry/exit points of methods. It also introduces async stack traces by passing context between threads, enabling deeper understanding of execution flow and exposing subtle performance bugs. Key technical challenges overcome include object pooling bottlenecks, global instrumentation event stream management, and multithreading issues with priority inversion, leading to the development of a lock-free ring buffer in Folly.
Jun 2014Improving Facebook on AndroidThis post details significant improvements made to the Facebook Android app focusing on performance, data efficiency, networking, and application size. Performance enhancements included deferring non-critical initializations and optimizing News Feed story loading, reducing start times by over 50%. Data efficiency was improved by adopting WebP for image compression (25-35% savings), requesting images at viewport resolution, and optimizing image caching, leading to a 50% reduction in data use. The networking stack was upgraded to OkHttp, enabling faster retries and SPDY protocol support, which reduced failed image loads by 90%. Application size was reduced by 65% through multi-APK uploads for different OSes and resolutions, with tools to track feature contributions to app size.
Mar 2013Under the Hood: Dalvik patch for Facebook for AndroidThis post details a critical patch applied to the Dalvik Virtual Machine for Facebook for Android to overcome method count limitations on older Android versions (Froyo and Gingerbread). The issue stemmed from the LinearAlloc buffer size being insufficient for the increased number of Java methods after a significant app rebuild. The solution involved a JNI extension to replace the existing buffer with a larger one and robust heuristics to handle variations in Android source code across different manufacturers, ultimately enabling the release of Facebook for Android 2.0 on a wider range of devices.
Dec 2012Under the Hood: Rebuilding Facebook for AndroidThis post details the native rewrite of Facebook for Android, focusing on performance improvements. Key contributions include: reducing garbage collection by minimizing allocations and deferring heavy tasks off the UI thread; implementing a custom, lightweight event bus to avoid reflection and object iterators; moving photos to the native heap using `inPurgeable` for reduced GC impact; and developing a custom ListView recycler to efficiently handle variable row heights and detach heavy content views. These efforts aimed to create a faster, smoother, and more reliable user experience.
Sep 2012Timed releases for mobile appsThis post introduces a date-driven release process for mobile apps, adapted from Facebook.com's daily release process. It emphasizes scheduling predictable release dates for testing, stabilization, and polishing to enable faster iteration, improve app performance, add features, and increase stability. The post highlights the unique challenges of mobile releases compared to web releases, such as higher stakes for app crashes and the inability to gradually roll out updates. It also discusses the importance of a high quality bar, targeted fixes, and code reuse across platforms. The goal is to balance rapid delivery with minimizing user disruption, aiming for 4-8 week release timelines.