Ext3 users - Aug 2002 - journal tuning

Hello,

  Is there some document about ext3 performance tuning and choosing
the right type and size of journal? Except RedHat's white paper.

  From what I read I understood that for typical operations
data=ordered is prefered. For the cases when there are many writes
not appending to files data=journal is the choice. And if I want to
get the most performance or in case where program is doing journaling
itself (Oracle) - data=writeback is the answer.

  But what about the size of the journal? The only clues I found is
that data=journal journal should be the size ~ 5 seconds worth of
writes to the disk. And that other types journals are typically
sized enough by default.

  System is Redhat 7.3.

  Thank you,

  Mindaugas

  P.S. What tool can show data flow per disk device? In my case
vmstat shows total but iostat and sar shows nothing about disk I/O.
SCSI - Compaq CISS Driver (v 2.4.30).

On Wed, Aug 21, 2002 at 09:55:57AM +0200, Mindaugas Riauba wrote:
>   From what I read I understood that for typical operations
> data=ordered is prefered. For the cases when there are many writes
> not appending to files data=journal is the choice. And if I want to
> get the most performance or in case where program is doing journaling
> itself (Oracle) - data=writeback is the answer.
My rationale is this:

I use data=journal for a Postfix queue. How much data can come in via
the 100MBit backbone in 5s (the commit interval):

100MBit/s * 5s = 500MBit

The journal should be about that size.
 
>   But what about the size of the journal? The only clues I found is
> that data=journal journal should be the size ~ 5 seconds worth of
> writes to the disk. And that other types journals are typically
> sized enough by default.
One could argue that the journal must be 5s * max transfer rate/s big.
How do I estimate that?

-- 
Ralf Hildebrandt (Im Auftrag des Referat V A)   Ralf.Hildebrandt@charite.de
Charite Campus Virchow-Klinikum                 Tel.  +49 (0)30-450 570-155
Referat V A - Kommunikationsnetze -             Fax.  +49 (0)30-450 570-916
Sysadmins don't go to hell; we're already doing our time in purgatory.

Hi,

On Wed, Aug 21, 2002 at 09:55:57AM +0200, Mindaugas Riauba wrote:
 
>   From what I read I understood that for typical operations
> data=ordered is prefered. For the cases when there are many writes
> not appending to files data=journal is the choice.
data=journal is really only expected to be a win if you have a lot of
synchronous IO (fsync and/or O_SYNC writes).  Otherwise it doesn't
really gain much, and you still pay the penalty of double-writes.

--Stephen