disk diagnostics

[Joerg Sonnenberger]

On Wed, Jul 26, 2006 at 08:06:57AM +0800, Bill Hacker wrote:
> >softupdates?  writing meta data with sync will be really slow.
> 
> No, not *that* slow, not even on K6-2-500 with 256 MB of SDRAM, where I 
> have done it on a production FreeBSD 4.8 web & mx box for donkey's years 
> (too small to hold a RAID). DFLY may not have focused on that area, but 
> should not be 2 or 3 orders of magnitude slower than 4.9/4.9 BSD.

I see typical disk troughput of 500-1000 transactions/second under load.
Let's assume 500 tps for now. Over 25min, that's 750k transactions. The
DragonFly tree alone has 42k inodes and the removal of a file needs at
least three IO transactions in sync mode (unlink from directory, update
of reference count in the inode, update of the block bitmap), resulting
in at least 120k transaction. Giving the low queuing and high seeking,
utilising a fourth part of the disk transaction rate doesn't sound that
bad. In fact, the result seems very likely when looking at the very
file systems comparisions.

Joerg