Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Why Intel added cache partitioning」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

Typical server utilization is between 10% and 50%. Google has demonstrated 90% utilization without impacting latency SLAs . Xkcd estimated that Google owns 2 million machines . If you estimate an amortized total cost of $4k per machine per year , that's $8 billion per year. With numbers like that, even small improvements have a large impact, and this isn't a small improvement.

How is it possible to get 2x to 9x better utilization on the same hardware? The low end of those typical utilization numbers comes from having a service with variable demand and fixed machine allocations. Say you have 100 machines dedicated to Jenkins. Those machines might be very busy when devs are active, but they might also have 2% utilization at 3am. Dynamic allocation (switching the machines to other work when they're not needed) can get a typical lat

测量时容易踩的坑

It's not obvious that this is possible. If both high and low priority workloads need to monopolize some shared resources like the last-level cache (LLC), memory bandwidth, disk bandwidth, or network bandwidth, there we're out of luck. With the exception of some specialized services, it's rare to max out disk or network. But what about caches and memory? It turns out that Ferdman et al. looked at this back in 2012 and found that typical server workloads don

For this graph, scale-out workloads are things like distributed key-value stores, MapReduce-like computations, web search, web serving, etc. SPECint(mcf) is a traditional workstation benchmark. “server” is old school server benchmarks like SPECweb and TPC . We can see that going from 4MB to 11MB of LLC has a small effect on typical datacenter workloads, but a significant effect on this traditional workstation benchmark.

对照之后能下的结论

Datacenter workloads operate on such large data sets that it's often impossible to fit the dataset in RAM on a single machine, let alone in cache, making a larger LLC not particularly useful. This was result was confirmed by Kanev et al.'s ISCA 2015 paper where they looked at workloads at Google . They also showed that memory bandwidth utilization is, on average, quite low.

You might think that the low bandwidth utilization is because the workloads are compute bound and don't have many memory accesses. However, when the authors looked at what the cores were doing, they found that a lot of time was spent stalled, waiting for cache/memory.

落到生产里的动作

Each row is a Google workload. When running these typical workloads, cores spend somewhere between 46% and 61% of their time blocked on cache/memory. It's curious that we have low cache hit rates, a lot of time stalled on cache/memory, and low bandwidth utilization. This is suggestive of workloads spending a lot of time waiting on memory accesses that have some kind of dependencies that prevent them from being executed independently.

LLCs for high-end server chips are between 12MB and 30MB, even though we only need 4MB to get 90% of the performance, and the 90%-ile utilization of bandwidth is 31%. This seems like a waste of resources. We have a lot of resources sitting idle, or not being used effectively. The good news is that, since we get such low utilization out of the shared resources on our chips, 工程上应当 be able to schedule multiple tasks on one machine without degrading performanc

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