在Go中友好的CPU缓存数据结构：相同算法下的10倍速度

Key Takeaways

关键要点

• Cache misses can slow down your code by 60x compared to L1 cache hits
• 与L1缓存命中相比，缓存未命中可能使你的代码慢60倍
• False sharing occurs when multiple cores update different variables in the same cache line
• 虚假共享发生在多个核心在同一缓存行中更新不同变量时
• Proper data structure padding can improve performance by 5-10x in specific scenarios
• 适当的数据结构填充可以在特定场景下提高5-10倍的性能
• Data-oriented design beats object-oriented for high-performance systems
• 数据导向设计在高性能系统中优于面向对象设计
• Always measure with benchmarks - cache effects are hardware-specific
• 始终使用基准进行测量 - 缓存效应是硬件特定的

Table of Contents

目录

L1 Cache:    4 cycles     (~1ns)      32KB
L2 Cache:    12 cycles    (~3ns)      256KB
L3 Cache:    40 cycles    (~10ns)     8MB
RAM:         200+ cycles  (~60ns)     32GB

Cache line size: 64 bytes (on x86_64)

Reading from RAM is approximately 60x slower than L1 cache. One cache miss equals 60 cache hits. This is why cache-friendly code can run significantly faster - often 5-10x in specific scenarios.

从RAM读取的速度大约比L1缓存慢60倍。一次缓存未命中等于60次缓存命中。这就是为什么缓存友好的代码可以显著更快运行 - 在特定场景中通常快5-10倍。

False Sharing: The Silent Killer

虚假共享：无声的杀手

False sharing occurs when multiple CPU cores modify different variables that happen to share the same cache line. This forces cache line invalidation across cores, causing significant performance degradation.

虚假共享发生在多个CPU核心修改不同变量时，这些变量恰好共享同一缓存行。这会导致跨核心的缓存行失效，从而造成显著的性能下降。

The problem is subtle: your variables might be logically independent, but if they're physically adjacent in memory (within 64 bytes), updating one invalidates the cache for all others on that line.

问题很微妙：你的变量可能在逻辑上是独立的，但如果它们在内存中物理相邻（在64字节内），更新一个会使该行上所有其他变量的缓存失效。

In our metrics collection system, we noticed 10x slower performance during high concurrency. The issue was multiple goroutines updating different counters that were packed in the same cache line.

在我们的指标收集系统中，我们注意到在高并发情况下性能下降了10倍。问题出在多个goroutine更新不同的计数器，这些计数器被打包在同一缓存行中。

Detection requires careful benchmarking with concurrent access patterns. The performance drop isn't visible in single-threaded tests, only under parallel load.

检测需要仔细的基...