Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Caches: LRU v. random」放到智能体、评测和线上系统里，真正要问的是：默认做法会不会系统性失败。下面用中文整理成可执行的工程笔记，去掉原站导航和无关链接。

问题怎么暴露

Once upon a time, my computer architecture professor mentioned that using a random eviction policy for caches really isn't so bad. That random eviction isn't bad can be surprising — if your cache fills up and you have to get rid of something, choosing the least recently used (LRU) is an obvious choice, since you're more likely to use something if you've used it recently. If you have a tight loop, LRU is going to be perfect as long as the loop fits in cache

In practice, on real workloads, random tends to do worse than other algorithms. But what if we take two random choices (2-random) and just use LRU between those two choices?

测量时容易踩的坑

Here are the relative miss rates we get for SPEC CPU 1 with a Sandy Bridge-like cache ( 8-way associative , 64k, 256k, and 2MB L1, L2, and L3 caches, respectively). These are ratios (algorithm miss rate : random miss rate); lower is better. Each cache uses the same policy at all levels of the cache.

Random and FIFO are both strictly worse than either LRU or 2-random. LRU and 2-random are pretty close, with LRU edging out 2-random for the smaller caches and 2-random edging out LRU for the larger caches.

对照之后能下的结论

To see if anything odd is going on in any individual benchmark, we can look at the raw results on each sub-benchmark. The L1, L2, and L3 miss rates are all plotted in the same column for each benchmark, below:

As we might expect, LRU does worse than 2-random when the miss rates are high, and better when the miss rates are low.

落到生产里的动作

At this point, it's not clear if 2-random is beating LRU in L3 cache miss rates because it does better when the caches are large or because it does better because it's the third level in a hierarchical cache. Since a cache line that's being actively used in L1 or L2 isn't touched in L3, an eviction can happen from the L3 (which forces an eviction of both the L1 and L2) since, as far as the L3 is concerned, that line hasn't been used recently. This makes it

To separate out the effects, let's look at the relative miss rates for a non-hierarchical (single level) vs. hierarchical caches at various sizes 2 . For the hierarchical cache, the L1 and L2 sizes are as above, 64k and 256k, and only the L3 cache size varies. Below, we've got the geometric means of the ratios 3 of how each policy does (over all SPEC sub-benchmarks, compared to random eviction). A possible downside to this metric is that if we have some ve

值得单独记下的观察

• Sizes are limited by dinero IV, which requires cache sizes to be a power of 2. [return]
• If you know of a cache simulator for some other domain that I can use, please let me know ! [return]

落地时建议先做的 5 件事

1. 用自己的真实负载测，而不是只用公开榜或厂商数字。
2. 把评测设计成能抓到失败模式：平均分好看但尾部崩溃，仍然算失败。
3. 智能体默认不会好好用测试；要写进流程，而不是写在口头规范里。
4. 性能和正确性都要有基线，改模型或改语言前后必须能对比。
5. 结论写成可回滚的决策：哪一版配置、哪一版评测集、谁签字。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，同样吃「先测量再扩面」这条纪律：技能、数字员工和网关都要有可复现评测，而不是只看一次演示通过。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源