Tackling Multi-CPU Performance Challenges with Baidu’s One-Click Btune

1  Multi-CPU Performance Challenges

Data centers now host a variety of CPUs (Intel, AMD, Ampere/ARM), making it difficult to ensure programs run at optimal performance across platforms. Traditional tuning requires deep hardware knowledge, extensive profiling tools, and expert analysis to identify true bottlenecks.

CPU differences appear at multiple layers:

Core level – instruction sets (AVX512 vs AVX256 vs Neon), SIMD support, frequency and hyper‑threading behavior.

Socket level – mesh vs multi‑die NUMA vs single‑die NUMA architectures, sub‑NUMA latency differences, L3 cache capacity and latency.

Interconnect level – CCIX (Ampere), xGMI (AMD), PCIe variations affect cross‑socket memory traffic.

Kernel level – CPM affinity, page‑cache placement, page‑table size, interrupt binding.

Runtime level – differing acceleration libraries (Intel MKL, AMD AOCL) and language runtimes (JDK, Python) have platform‑specific optimizations.

Application level – code paths may behave differently when migrated between x86 and ARM/AMPERE.

These challenges raise the difficulty of scaling performance tuning across heterogeneous fleets.

2  Btune One‑Click Optimization Design

Btune automates the four classic tuning steps: metric collection, bottleneck identification, performance optimization, and SLA verification. It integrates more than ten profiling tools covering over a hundred metrics across four dimensions (hardware, kernel, runtime, application).

Key analysis methods include:

USE – evaluates resource utilization, saturation, and errors.

TSA – time‑slice analysis to find longest‑running phases.

TMA – CPU‑resource bound analysis.

Btune builds a “bottleneck analysis tree” that traverses from high‑level resource distribution down to specific leaf‑node causes (e.g., TLB miss, sub‑NUMA latency, missing huge pages). The system then matches each leaf node with expert‑curated optimization recommendations stored in a knowledge base.

Typical workflow:

Select instance and process, click “One‑Click Analyze”.

Btune generates a concise analysis summary (bottleneck points and suggested actions) and a detailed report (system config, thread model, hot‑spot functions, etc.).

Btune is free on Baidu Intelligent Cloud BBC and BCC compute instances, and an accelerated version (BtuneAK) will be released soon.

3  Baidu Cloud Optimization Practice

Three real‑world cases illustrate Btune’s impact:

Search subsystem latency reduced 3.9‑4.6% by enabling transparent huge pages to fix a TLB bottleneck.

BRPC‑based ranking service lowered CPU utilization by 25.8% after switching Bthread scheduling mode.

Storage service cut average request latency by 17% and 99th‑percentile latency by 11.7% by enforcing node‑local disk affinity.

Across these examples, Btune’s automated analysis and expert recommendations closed the loop from bottleneck detection to concrete performance gains.