AW: AW: AW: [OpenAFS] Severe Performance Issue with SCSI!

[Rubino Geiß]

Thanks,

oh, I guess despite my long posting I forgot some measures I have
already taken. A few week ago I thought the same as you did, but now,
all of it has been tested: no significant difference!

In detail:
	- I tested ext3 / ext2 and reiserfs for server partitions
	- I tested especially every mode of ext3 in combination with
different block size and journal file size. All I get was an
insignificant change +- 10 .. 30 % 
	- In every case heavy disk io was produced...
	- If I put the journal on a memory file system and put ext3 to
data=journal, then I could gain most significant speed. But in our test
case it would be sill slower than a fast IDE disk.

	- Of course I considered SCSI hardware disk caches, fist: they
are on.

Anyway, thank you,
Rubino

-----Ursprüngliche Nachricht-----
Von: Helge Bahmann [mailto:bahmann@math.tu-freiberg.de] 
Gesendet: Montag, 5. August 2002 11:43
An: Rubino Geiß
Cc: openafs-info@openafs.org
Betreff: Re: AW: AW: [OpenAFS] Severe Performance Issue with SCSI!


What filesystem are you using on the server? From your description it
appears that the hard disks are seeking like mad to honour the
transaction integrity of the filesystem under synchronous writes; if it
is ext3, then data=journal is your friend for small files

It appears that your IDE disks are running with write caches enabled and
are ignoring cache flush commands, allowing them to reorder and batch
writes