Having had a look at the results /proc/stat is handing out for disk io,
Linux tends to be pretty closed compared to other OS.
There are 4 numbers per entry, and neither the entries reveal good
information about the harddisks nor are the numbers really meaningful
in some configurations.
E.g.: If you have 4 IDE-disks, they will be assigned well, but if
you mix SCSI- and IDE-hds, things won't work nice, and if you are using
the md-driver, it will really get mixed up.
I used to work at a Sinix-machine, which provided the following entries:
- Block reads/write
- number of requests in the waiting queue when a new request
arrived.
- Waiting time of the requests in the queue
- Time for completion of the requests
Everything is divided up for each partion (instead of harddisk), so
you can analyse how to move single file systems to different disks
to distribute the load between the disks.
I started developing a patch for Linux 2.0.27 (its really not that
far developed that I'd like to post it to anybody), but today a
friend told me that performance measurement of harddisks is somewhere
defined in the POSIX (the? which?)-standart.
Now I'd like to know:
- Is anybody working on something like that?
- Does anybody have a reference or might mail me the part where
this performance measuring is defined?
- Does anybody have an idea if there should other statistics than
those above be measured by Linux?
The reason I started to develope the patch for 2.0.27 is that I
have to do production work at the ISP I'm working for and that
I need the information for production machines. I'll port it
to 2.1.x once its ready.
Sorry if something similar should already be in the 2.1.x-kernel,
I'm a bit out of the most recent developements.
Greetings,
Georg v.Zezschwitz