Alibaba's X86 CPU Customization Journey for Double 11: From Specification Selection to Feature Customization

Introduction: Alibaba proudly recounts its 2017 Double 11 achievements—32.5 million transactions per second, 42 million database operations per second, and 2.56 million payments per second—highlighting the smoothness of the event and announcing a series of technical shares on the underlying infrastructure.

Background: The success of Double 11 depends on a massive coordinated system from top‑level applications down to server clusters, with custom X86 CPUs playing a key role. As Moore's law slows and Intel’s Tick‑Tock model shifts to a Process‑Architecture‑Optimization (P‑A‑O) cycle, traditional performance gains are insufficient for Alibaba’s growing compute demand.

Horizontal scaling can partially meet demand, but Total Cost of Ownership (TCO) limits its scale. Optimizing TCO involves either increasing the effective service capability per unit of CapEx or reducing the OpEx associated with that CapEx, while keeping total compute, storage, and I/O resources constant.

X86 CPU Customization Path: Facing Moore’s law slowdown and TCO pressures, Alibaba focuses on aligning CPU architecture with its specific workload characteristics rather than relying solely on generic processors.

Specification Selection: The first step is choosing from vendor‑provided SKUs based on Alibaba’s TCO model. These SKUs offer coarse‑grained dimensions such as core count, frequency, and cache size. The ideal configuration often differs significantly from the available SKUs, leading to performance and efficiency gaps.

Specification Customization: Alibaba then requests the CPU vendor to create new SKU configurations tailored to its workload, narrowing the gap between the ideal and actual configurations and achieving noticeable energy‑efficiency improvements.

Feature Customization: Going beyond specification tweaks, Alibaba introduces new configuration dimensions, often realized through micro‑code patches that add or modify instructions to dramatically boost performance for its specific workloads.

Outlook: The CPU customization program has already delivered substantial benefits, ensuring reliable compute power for Alibaba’s services, especially during Double 11. Future work will deepen the customization process, refine the loop from workload to CPU architecture and back, and continue iterating to provide sustained compute upgrades.