Re: [PATCH] xfs: add per-mount read/write I/O completion counters

[Eric Peterson <linuxinstalled@gmail.com>]

Eric,

I have an application that sends IO through several different layers.
In my case:
Ethernet adapter -> NFS -> XFS -> Block device -> SSD

I already have visibility of qdepth/RT and IOPS at the ethernet
adapter, block dev, and SSD, but I'm missing information about what is
going on in the middle.

I wanted a "light touch" way of measuring what is going on in the middle
so that I could have hard data about the performance impact of my
changes.
As part of that "light touch", I wanted to avoid a method that would
artificially inflate metrics due to the increased load from
instrumentation.
In addition, my approach was to follow the existing code structure of
counters that were already present.

Regarding Dave's points about edge cases he illustrated, I am in
agreement with them on the technical basis.
If precision is decided to be more important than overhead, then this is
not the right course of action.
However, I disagree that the edge cases make this a "useless" way of
taking measurements.
As I shared, there are still several applications that would benefit
from this.

Regarding whether these counters get merged or not, I already have them
present in my kernel builds.
My personal needs are met.
I chose to submit a patch because I believed others could also find
benefit from them.
If the cost of the stat was cheap enough, and given the right
understanding of the caveats of this method, it was worth sharing.

-Other Eric