Is there a way to show linux buffer cache misses?

Question

I am trying to measure the effects of adding memory to a LAMP server.

How can I find which processes try to read from the Linux buffer cache, but miss and read from disk instead?

Answer 1

SystemTap is one of the best ways to do this, but fair warning it's difficult to get a great answer. The kernel simply doesn't provide this data directly. You have to infer it based on how many times the system requested a read and how many times a disk was read from. Usually they line up fairly well and you can attribute the difference to the VFS cache, but not always. One problem is LVM- LVM is a "block device", but so is the underlying disk(s), so if you're not careful it's easy to double-count the disk reads.

A while back I took a stab at it and wrote this:

https://sourceware.org/systemtap/wiki/WSCacheHitRate

I do not claim that it is perfect, but it works better than nothing, and usually generates reasonable output as long as the environment is fairly "normal". It does attempt to account for LVM in a fairly crude way.