How Baidu Optimized Low‑End Device Startup Performance: A Deep Dive

Background

With the mobile internet user base reaching its peak, growth has slowed and companies are shifting focus from acquiring new users to retaining existing ones. Low‑end devices, which make up a significant portion of the retained user base, have become a critical performance bottleneck for apps such as Baidu App.

Goals and Value

The project aims to improve user experience on low‑end phones by optimizing startup performance, establishing a performance‑driven design, creating reliable problem‑location mechanisms, and building a full performance‑optimization loop.

Key Challenges

Complex startup flow leads to severe black‑screen and stutter issues on low‑end devices.

Lack of an overall scheduling mechanism for numerous pre‑load tasks.

High cost of locating performance problems with existing tools.

Insufficient anti‑degradation mechanisms, relying only on partial online metrics and user feedback.

Approach Overview

The solution is split into three sub‑directions: observability facilities, core infrastructure, and business‑level optimizations.

1. Observability Facilities

Define low‑end devices using a scoring system (static + dynamic scores) where devices with PV/UV share <15% are classified as low‑end.

Build startup performance metrics, primarily Time‑to‑Interactive (TTI), adapted from web metrics (FCP, FID) to mobile.

Establish online and offline anti‑degradation pipelines to detect regressions early.

2. Core Infrastructure

High‑Efficiency Performance Tools : Developed a Trace tool (ASM‑based instrumentation + Perfetto) and a Hook tool (Epic/XHook based) for visualized, automated analysis of lock contention, long tasks, I/O, and thread issues.

High‑Performance KV Storage (UniKV) : Replaced Android SharedPreferences with a multi‑process, non‑blocking KV store supporting strong typing, data migration, and reduced thread usage, achieving up to 95% faster first‑read performance on low‑end phones.

Lock Optimization : Refactored ABTest component to eliminate synchronized blocks, introduce lock‑free reads, and use JSON/PB formats for cached data, cutting TTI by >200 ms.

Scheduling Framework : Designed a task scheduler that accepts business‑defined tasks, gathers device, behavior, and scene profiles, and provides personalized, tiered, and priority‑based execution, enabling fast integration of pre‑load optimizations and parallel rendering of splash screens and home pages.

3. Business Optimization

Problem discovery via Trace/Hook tools (offline) and online metrics/experiments (online).

Collaboration with product teams to prioritize fixes, schedule releases, and validate impact.

Verification of optimizations through regression testing and monitoring of both performance and business KPIs.

Results

The combined effort delivered measurable improvements:

Cold‑start TTI on Android reduced by >50 % and on iOS by >40 %.

First‑launch TTI on Android improved by ~40 % and on iOS by ~30 %.

ABTest lock contention eliminated, yielding a 200 ms+ TTI gain.

UniKV adoption across >100 files, reducing thread count and eliminating ANR cases.

Future Outlook

Continued investment will expand beyond startup to the full app lifecycle, focusing on view‑system smoothness, automated anti‑degradation pipelines, and broader adoption of the scheduling framework across more scenarios such as splash screens, external launches, and core feed rendering.

References

Web FID: https://web.dev/fid/

Epic (hook framework): https://github.com/tiann/epic

xHook: https://github.com/iqiyi/xHook

Perfetto docs: https://perfetto.dev/docs/

MMKV: https://github.com/Tencent/MMKV

Booster: https://github.com/didi/booster