Very low disk performance on 5.x

[Robert Watson]

On Mon, 2 May 2005, Poul-Henning Kamp wrote:

>> I'm quite willing to test and optimise things but so far no one has
>> had any concrete suggestions on that to try.
>
> First thing I heard about this was a few hours ago.  (Admittedly, my 
> email has been in a sucky state last week, so that is probably my own 
> fault).

This is because I resent a message on the -performance list to you and 
Scott Long, and Soren because I was concerned that the thread was 
identifying a reproduceable problem for several people, but no one had 
really bit in to try and diagnose it :-).

It seems likely we're hurting at a couple of levels, and it would be good 
to identify what's appearing at what level.  I think comparing dd results 
on the total disk device directly as a starting point is a good place to 
begin.  I'm not sure if we've seen Linux and FreeBSD dmsg output yet, but 
if nothing else it would be good to confirm if the drivers on both systems 
negotiate the same level of throughput to each drive.

The next thing that would be quite nice to measure is the rate of I/O 
transactions per second we can get to the disk using the disk device 
directly, with a minimal transaction size.  I have a vague recollection 
that you have to be careful in Linux because their character device nodes 
for disk devices are buffered, and you really want unbuffered I/O.

After transactions/sec, broad sweeps across well-understood areas of the 
disk would be a good thing to do, again using the device node directly.

Isolating the impact of the file system will be important...

Robert N M Watson